Results in Engineering 25 (2025) 103676

Contents lists available at ScienceDirect

Results in Engineering
journal homepage: www.sciencedirect.com/journal/results-in-engineering

Review article

Review of machine learning applications in additive manufacturing

Sirajudeen Inayathullah , Raviteja Buddala *
Department of Manufacturing Engineering, School of Mechanical Engineering, Vellore Institute of Technology, Vellore, Tamil Nadu, 632014, India

A R T I C L E I N F O A B S T R A C T

Keywords: The necessity to produce intricate components results in considerable progress in manufacturing methods. Ad-
Additive manufacturing ditive manufacturing (AM) is a disruptive technology that allows intricate and custom-tailored components to be
3D printing fabricated with great precision and efficiency. It is applied in advanced sectors like aerospace, healthcare,
Artificial Intelligence
automotive industries, and it starts having their interest in many other areas. Machine learning (ML) has become
Deep learning
Machine learning
a powerful tool for overcoming problems in AM, offering process efficiency, defect detection, quality assurance,
Process optimization and predictive modelling of mechanical properties. This review discusses how ML transforms AM by providing
Defect detection design evaluation, process optimization, and production control innovation. The approach taken in the study is
systematic, examining the current literature and case studies of ML application to AM. Hybrid data collection
techniques that combine machine settings with physics aware features and yield robust predictive models are the
focus. Additionally, the review evaluates various ML algorithms used to predict mechanical properties, optimize
process parameters, and characterize AM processes. The measurements indicate groundbreaking improvements
in ML powered solutions, like process monitoring in real time, automatic parameter adaptation, and defect
mitigation that offer greater accuracy, ease, and reliability in AM. Yet, data scarcity, computational challenges
and a gap between research and industrial applications of ML exist. To realize the full potential of ML in AM it is
critical to address these challenges. It closes with the identification of promising research directions including
standardization of data improvement, developing new advanced ML algorithms, and building an interdisci-
plinary research effort to spur additional progress in this field.

(continued )
List of abbreviations
DQN Deep Q Networks
Abbreviation Meaning PPO Proximal Policy Optimization
AM Additive manufacturing GP Gaussian Processes
ML Machine learning GANs Generative Adversarial Networks
AI Artificial Intelligence ANN Artificial neural network
DL Deep Learning ANN-GA Artificial neural network-Genetic algorithm model
CAD Computer Aided Design CNN Convolutional Neural Networks
STL Standard Tessellation Language KNN k-nearest neighbor algorithm
SLM Selective laser melting SOM Self-organizing map algorithm
FFF Fused filament fabrication VAEs Variational Autoencoders
SLA Stereolithography CAEs Contractive Autoencoder
LENS Laser-fabricated net shaping PID Proportional-Integral-Derivative
PRISMA Preferred Reporting Items for Systematic Reviews and Meta- MAML Model-Agnostic Meta-Learning
Analyses PSO Particle Swarm Optimization
TO Topology optimized PINNs Physics-Informed Neural Networks
RL Reinforcement Learning LSTM Long short-term memory networks
RF Random Forest AE Auto encoder
DT Decision tree DfAM Design for Additive Manufacturing
TL Transfer learning SVMs Support Vector Machines
LR logistic regression SVR Support Vector Regression

(continued on next column) (continued on next page)

* Corresponding author.
E-mail address: [email protected] (R. Buddala).

https://doi.org/10.1016/j.rineng.2024.103676
Received 23 August 2024; Received in revised form 25 November 2024; Accepted 5 December 2024
Available online 8 December 2024
2590-1230/© 2024 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-
nc-nd/4.0/).
S. Inayathullah and R. Buddala Results in Engineering 25 (2025) 103676

(continued ) cannot be manufactured without carefully following each of these

SVMR Support vector machine regression stages. Monitoring and inspecting processes and components extensively
GNN Graph neural network have been the focus of recent attempts to decrease AM part variability.
CNC Computer numerical control Progress in sensor technology, sensor fusion, and data gathering meth-
PDM Predictive maintenance
PHM10 Prognostics and Health Management 2010 data set
odologies has resulted in a remarkable surge in AM data, covering
RUL Remaining Useful Life numerous elements of big data. The many categories of data produced
SSM State space model right through from conception to the final output conversion cycle are
DDIM Denoising diffusion implicit models presenting fresh prospects for knowledge exploration inside AM pro-
DNA Deoxyribonucleic acid
cedures. AM has become a data-rich industry with limited knowledge. It
LB-PBF Laser-based powder bed fusion
XGBoost Extreme Gradient Boosting may be difficult to gain insights from copious AM data. Despite gains in
XAI Explainable Artificial Intelligence measurement science and more AM lifecycle datasets, there is a dearth
MLP Multi-Layer Perceptron of scientific understanding to fully define the relationships between AM
MPC Model predictive process control materials, structure, process, geometry and quality [3–5].
AutoML Automated Machine Learning
GRB Gamma-Ray Bursts
Sophisticated computational and analytical approaches are required
PCA Principal Component Analysis to handle the highly detailed and intricate datasets. Recent de-
FPCA Functional principal component analysis velopments in ML have dramatic prospects for addressing this gap.ML
PLA Polylactic Acid approaches have the capability to uncover implicit (previously un-
DTs Digital Twins
known) knowledge and detect correlations in extensive production
HME Hot melt extrusion
CTGAN Conditional Tabular Generative Adversarial Network datasets, converting vast amounts of data into practical and enlightening
HSI Hyperspectral imaging information [6,7]. This study examines the distinctive function of ma-
PINN Physics-Informed Neural Networks chine learning in tackling the issues associated with AM. This study
SurpriseAF- Surprise-guided sequential learning framework systematically reviews existing literature to analyze how ML approaches
BO
might enhance different phases of the AM lifecycle. Specifically, it em-
ResNet Residual Network
ConvNeXt A pure convolutional model phasizes how ML facilitates enhanced process monitoring, real-time
ViT Vision Transformers parameter modifications, and predictive analytics, so augmenting the
SWIN Shifted window reliability and performance of AM processes. Similarly, fabrication
Edge AI Edge artificial intelligence
processes have been advanced due to the growing demand for complex
NSAI Neuro-symbolic AI
GCM Generative Convergence Model components, especially in multi material printing, which enables crea-
MOGOA Multi-objective Grasshopper Optimization Algorithm tion of different material properties within the same build to enhance
LGBM Light Gradient-Boosting Machine functionality and performance [8]. Direct energy deposition and powder
ERT Electrical resistance tomography bed fusion innovation is also supportive in the creation of complex ge-
LOOCV Leave One Out Cross Validation
ometries with minimum waste while maintaining maximum accuracy
[9]. Moreover, with the addition of advanced modeling tools and AI
generated design variants, manufacturers can save weight and optimize
structural efficiency to meet demands of aerospace and automotive in-
1. Introduction dustry [10]. Along with the capacity to analyze and optimize AM pro-
cesses, these developments could change the manufacturing landscape.
AM is a progressive and widely used favored production technique
that represents the groundbreaking advancements of the contemporary 1.1. Additive manufacturing
manufacturing sector. Not like more typical manufacturing methods,
which in most cases use material removal, as in milling or shaping or as AM, an age-old technique is now being used to create avant-garde
with casting, AM produces components by accumulating materials layer products in the contemporary world of elegant new-age manufac-
by layer using data from a three-dimensional model. AM provides a turers. This is a cool technology that builds components by stacking
competitive edge compared to conventional manufacturing methods. It layers of material with 3D model data (Fig. 1). The initial step involves
allows for the production of customized goods with intricate shapes and generating a 3D model of the product’s features with Computer Aided
specific material qualities, even in small quantities. This is achieved in a Design (CAD) modeling software, adhering to the provided re-
cost-effective and time-efficient manner [1]. However, due to this quirements. After the CAD model is created, it is divided into sections
versatility and efficiency, AM hold great value in industries that require using parallel planes that are the same thickness as each layer. Slices will
rapid prototyping or complex geometries. AM allows for greater flexi- be created by dividing these three-dimensional models into smaller,
bility in design with geometries far more complex than equivalent two-dimensional items. Basically, what was once a complex three-
quantities of traditional materials and less waste generated in the pro- dimensional difficulty is now only a collection of two-dimensional
cess. It also speeds prototyping without tooling costs, which traditional challenges. The Standard Tessellation Language (STL) format is used
methods are more material and time costly for custom or complex de- to transmit these compact two-dimensional data, which originate from
signs. AM is therefore advantageous for faster, more accurate produc- tessellating a 3-dimensional geometric layout. In a CAD model, the
tion. Due to the ongoing technical advancements, AM has progressed surface normal and the coordinates of the vertices are kept, and a
beyond its initial role of creating prototypes and is now capable of tessellation sequence of triangles is used to piecewise approximate each
generating functional metallic components for many applications, surface.
including automotive, defense, biomedical, and aerospace industries The ability to quickly prototype, iterate designs and increase the
[2]. testability of products through 3D printing has a massive impact on both
Despite developments in the AM sector, consistent component product design review and quality assurance. CAD is already faster than
quality and process dependability remain a barrier. The primary reason developing a physical prototype, but it allows designers to quickly test,
for this condition is because the AM process forms both a part’s shape form, fit and function. Test, input from users and continuing to iterate
and material characteristics simultaneously. AM parts are created by a helps the design process move at a faster clip while also kickstarting
multi-stage process involving extensive design, material, and process creativity. Moreover, 3D printing can achieve complex shapes with
interactions. There are five basic steps: process design, process planning, tailor-made details not feasible by conventional manufacturing [11,12].
construction, postprocessing and testing & validation. A certified part The integration of sensors and the in-situ monitoring using 3D printing

2
S. Inayathullah and R. Buddala Results in Engineering 25 (2025) 103676

Fig. 1. AM Process.

to identify anomalies or deviations from specifications during a print the processing of such a complex and high-dimensional data. Promising
process is noted as well [1,13]. When the system incorporates its own new breakthroughs in the science of ML have the ability to convert AM
quality controls, it generates few customer complaints and instills rising data into meaningful knowledge [20,21].
customer satisfaction in other ways that might also save on waste of
production materials. 1.2. Overview of ML techniques
AM offers a distinct advantage over conventional production tech-
niques. Producing bespoke goods with complex designs and one-of-a- ML is an AI technology that enables data analysis and decision-
kind material qualities has never been easier or faster. Parts made of making without the assistance of specially programmed software. Aca-
metal, polymers, and ceramics may be produced using a number of AM demic disciplines such as medical diagnostics, smart manufacturing,
subcategories, which have emerged as a consequence of the technol- autonomous driving, object identification, natural language analysis,
ogy’s meteoric rise. Laser-fabricated net shaping (LENS), selective laser and material property have all been on the rise lately [22]. The changes
melting (SLM), fused filament fabrication (FFF), and stereolithography can either improve the operation of completely new systems or the
(SLA) are a few of the AM fabrication techniques that have been current ones in the 1st version. Monitored learning is an ML activity that
developed to print actual, functional parts from a variety of material uses labelled training data to create a function. Training data is a
types and shapes. The continuous development of AM technology has collection of the training instances. In supervised learning, every
allowed it to transcend beyond the fabrication of prototypes and into the example consists of a vector as input and valued output, often referred to
production of metallic components for end-use in several industries, as the supervisory signal. To map new instances, a supervised learning
such as aerospace, defence, biomedical, and automotive, to name a few. system uses an inferred function it builds from training data [23,24]. ML
The AM industry has come a long way, but there’s still a long way to go algorithms are usually classified into semi-supervised, reinforcement
before component quality and process dependability are consistently learning, supervised learning, and unsupervised learning (Fig. 3).
met [4,14–16].
The primary reason for this condition is because the AM process 1.2.1. Supervised learning
forms both a part’s geometry and material characteristics simulta- Supervised learning entails the process of instructing an algorithm
neously. Fig. 2. illustrates the AM processes using ML techniques. The using a dataset, where each data point is accompanied by a corre-
fabrication of any additively manufactured component necessitates sponding label. This label indicates the specific group to which the ex-
intricate interactions between design, material, and process throughout ercise point fits. Once the data clusters have been defined, supervised
a multi-stage, complex procedure comprising of five main stages: design, algorithms will search for the decision boundaries that divided them.
process plan, creation, post-process and testing & validation [5,17–19]. Supervised learning techniques establish a model that represents the
AM has evolved into an industrial sector that is both data-rich and correlation among the input characteristics and labelled outputs.
knowledge-scarce. Extracting insights from the vast amounts of avail- Therefore, it has the capability to anticipate input characteristics for
able AM data may be a time-consuming process. Scientific understand- intended results (Fig. 2). Supervised learning is applied very widely in
ing of the relationships between AM materials, geometry, process, AM, where the input parameters-including speed, temperature, and the
structure, property, and performance is limited, even though measure- material type-are mapped to the desired outputs, such as strength, sur-
ment science and the amount of AM lifetime datasets are both on the face finish, or defect rate.
rise. Sophisticated computational and analytical methods are needed for

Fig. 2. AM processes using ML technique.

3
S. Inayathullah and R. Buddala Results in Engineering 25 (2025) 103676

Fig. 3. The approaches of ML.

1.2.2. Unsupervised learning sophisticated application-specific algorithms toward AM, including

Unsupervised algorithms don’t use a labeled dataset like the super- real-time control, defect detection, and optimization over
vised because unsupervised is not about predictions (predictive multi-dimensional data. Advanced ML algorithms thus include, for
modeling). Unsupervised learning receives unlabeled data as input and instance, RL, for dynamic and feedback-driven adjustments; Bayesian
creates its own clusters using the inherent characteristics of the supplied models for predictions with limited data; and generative models, such as
objects to automatically sort it. Therefore, these models are used to GANs, for simulating realistic patterns in designs and defects on the
understand the latent or unknown relationships between data (Fig. 3). parts. It includes transfer learning, meta-learning, deep learning archi-
Learning without supervision deals with uncovering latent patterns in tectures and evolutionary algorithms as well hybrid models that
the unlabeled data. In AM, this is useful for large volumes of process data improve capability of ML in AM generally by improving flexibility, ac-
analysis and underlying trends identification. curacy or efficiency across different kinds of AM processes [25]. This
wide variety of traditional and advanced ML approaches including en-
1.2.3. Semi-supervised learning ablers the creation of more robust, adaptable AM systems that are
A combination of unsupervised and supervised methods is what applied in many tasks such as quality control, parameter optimization or
makes up semi-supervised learning. Semi-supervised learning models multi-objective decision making. Summary on advanced or complicated
combine labelled and unlabelled data because it is infeasible to label ML techniques utilized in AM are listed in Table 1.
millions of instances, leading to high human costs. These types of models
that may have some labelled data but the majority are unlabelled and
utilise a small amount of supervised learning are referred to as semi-
supervised learning (Fig. 3). These models are less experienced in su-
Table 1
pervised learning (expensive and difficult to teach). Semi-supervised Advanced ML Algorithms used in AM.
learning is helpful in AM whenever labelled datasets (like defect data)
Algorithm Specific Algorithms Application in AM
are scarce but there is ample sensor or process unlabelled data available.
category

Reinforcement Q-Learning Real-time optimization of

1.2.4. Reinforcement learning
Learning machine settings throughout the
RL techniques use training data that can only indicate whether or not printing process
they are accurate, in contrast to supervised learning, which uses labelled Deep Q Networks (DQN) Acquiring optimum
data. They acquire "good" behaviour through recurrent interactions with configurations from high-
their environment. RL is a learning method that operates similarly to dimensional datasets
Proximal Policy Refining continuous control
supervised learning, except instead of relying on a vast amount of
Optimization (PPO) tasks (e.g., print head motion)
labelled data, the model interacts with the environment and receives Bayesian Models Bayesian Optimization Optimization of process
positive rewards or negative punishments. This feedback strengthens the parameters with minimum
behavior of the model, resulting in its designation as RL algorithm experimentation
Gaussian Processes (GP) Surrogate modelling for
frequently utilise the words exploitation and exploration (Fig. 3). In
forecasting performance or
contrast to exploration, which is doing something new, exploitation mechanical characteristics
involves doing something that will provide the most potential benefit. Generative Generative Adversarial Creating synthetic component
By integrating these two approaches, the model may systematically Models Networks (GANs) designs and forecasting flaws
acquire knowledge about the surroundings and recognize inputs that Variational Autoencoders Modelling variability in AM
(VAEs) processes and modelling defect
yield favorable rewards, ultimately leading to optimal solutions.
formation
Transfer Learning Pretrained CNN Modifying flaw detection models
(Convolutional Neural for various materials
1.3. Advanced ML algorithms in AM Networks) models
Meta-Learning Model-Agnostic Meta- Rapid adaption of models to
Learning (MAML) novel AM settings or machinery
ML technology offers many algorithms for solving diverse chal- DL Architectures CNN Image analysis for defect
lenges. These algorithms have a common purpose of learning and pre- identification and component
dicting future states of a process based on existing data. Researchers quality evaluation
have used a wide range of ML methods. Neural networks (NNs), genetic Particle Swarm Identifying ideal parameters for
Optimization (PSO) efficiency or energy utilization
algorithms (GAs), response surface methodology (RSMs), fuzzy infer-
Hybrid Models Physics-Informed Neural Integrating domain expertise
ence systems (FISs), and regression modelling are among the most often Networks (PINNs) with data-driven algorithms for
used approaches. Support vector machines (SVMs), simulated annealing precise predictions
(SA), finite element analysis (FEA) and more are examples of less used Hybrid Bayesian Networks Integrating empirical data with
approaches [8]. Beyond these fundamental methods, newer break- expert knowledge in decision-
making
throughs in ML have come along with the introduction of more

4
S. Inayathullah and R. Buddala Results in Engineering 25 (2025) 103676

1.3.1. Bayesian methods clearly stated research issue, a systematic review will collect, evaluate,
Bayesian approaches play a critical role in the optimization process and synthesize accessible empirical data and research using predefined
of AM, as they deal with uncertainty, thereby improving decision- criteria. The study’s research question is as follows: What are the
making. Bayesian optimization is repeatedly used for optimizing most different ML approaches used in AM, and what are the current de-
of the parameters that define a process, such as print speed, temperature, velopments in these applications? To ensure the practice of high rigor
and layer height, by learning from previous outcomes and repeatedly and transparency, this systematic review adhered strictly to the PRISMA
adjusting predictions to minimize the number of trials to be taken to (Preferred Reporting Items for Systematic Reviews and Meta-Analyses)
obtain optimum settings, thereupon saving time and resources. GP are a standards. It therefore involved many stages, including the search for
Bayesian methodology that provide probabilistic forecasts for mechan- relevant studies, screening based on eligibility criteria, and summarizing
ical properties or failure rates. These enable the dependability evalua- findings. This review assessed previous research in many ML method-
tion of printed parts under process variations. These technologies deliver ologies utilized by the AM industry for process optimization, quality
data-driven optimization-that is more effective in complex AM control, and other purposes.
workflows. The literature review comprises all papers published between 2014
and 2024, selected using the criteria indicated in Table 2. The Scopus
1.3.2. Generative models database was chosen as the principal source because of its extensive
Generative methods are transformational for AM since they allow for interdisciplinary indexing. The relevance and quality of the studies
the creation of new geometries and optimize complex geometries. GANs reviewed were assured through a set of inclusion and exclusion criteria.
are applied for the generation of new part designs; with such capability, The inclusion criteria involved discussions on the ML application to AM,
new structures can be explored in order to achieve an enhanced per- studies addressing life cycle stages, including design, process planning,
formance. VAEs are used for creating realistic variations of part geom- build, post-processing, and quality assurance, as well as peer-review
etries especially when data are limited, hence expanding possibilities of articles published in the English language. On the contrary, the exclu-
design. Such models have enhanced optimization of designs for perfor- sion criteria left out non peer reviewed articles, other studies which do
mance and material efficiency, and AM innovation has resulted from the not focus on ML in the context of AM and articles published outside the
automation and enhancement of the design process based on learned specified time frame (Fig. 4). Every publication underwent a compre-
patterns. hensive analysis, from which the subsequent information was taken:
publication year, paper title, specific AM process (e.g., fused deposition
1.3.3. Transfer learning modeling), study purpose (quality assurance, process chains, sched-
TL approaches enable the transfer of knowledge between domains uling), input variables, output variables, and the ML algorithm(s)
and, as a consequence, reduce data requirements for model training. employed. The investigation’s objectives are achieved through the uti-
Transfer learning: Pretrained models, mostly CNNs in AM can be fine- lization of this data. The subsequent section delves into the findings and
tunes to detect anomalies or investigate material/machine usage pat- discoveries of the literature assessment. The findings were examined
terns offering better generalization capabilities and reduced computa- over different AM processes and applications of ML to make them
tional burden. This enables rapid deployment of ML solutions with generalizable. For example, research falls under several disciplines,
different asset management platforms. including engineering, computer science, materials science, medicine,
etc. Figs. 5 through 8 offer a glimpse into publication trends, important
1.3.4. Meta learning contributors, and research worldwide.
Meta-learning is known as the process of learning mechanisms and Fig. 5 illustrates the number of Scopus articles associated with ML
strategies of the learning process itself, or "learning to learn." This papers in the AM field from 2014 to 2024. It is clearly evident that the
teaches models to adapt rapidly to new tasks or AM environment field is evolving year by year. Fig. 6 illustrates the diverse author con-
changes. Algorithms like MAML allow different materials, geometries, tributions to the Scopus publications associated with ML papers in the
or types of AM machines the ability to adapt in the models and therefore AM field, spanning from 2014 to 2024. Abdul W Basit has contributed
improve flexibility in dynamic production environments. the greatest number of publications, followed by Pedro Cabalar, Ales-
sandra Caggiano, and several other notable researchers. From 2014 to
1.3.5. Deep learning architectures 2024 (Fig. 7), there will be a large number of ML articles published
DL particularly CNNs and RNNs, enable even complex processing of under the AM field across a variety of broad research areas (Engineering,
data within AM. CNNs are very valuable in real-time defect detection Computer Science, Materials Science, Physics and Astronomy, Chemical
and quality inspection due to their capabilities in handling visual data. Engineering, Medicine, Biochemistry, Genetics and Molecular Biology,
The capability of the RNNs in real-time processing of time-series data, Energy, Mathematics, Pharmacology, Toxicology, and Pharmaceutics,
such as monitoring the material flow and temperature monitoring is among others), according to Scopus. The Engineering area has the
responsible for better observation of the process and stability. greatest number of articles, followed by computer science and materials
science. Fig. 8 shows the number of Scopus journals associated with ML
1.3.6. Hybrid models
Hybrid models integrate physics-based concepts with data-driven ML
techniques to enhance predictive accuracy. PINNs integrate physical Table 2
rules, such as thermodynamics, into the model, rendering them very The criteria used to choose the available research for the review.
effective for precise predictions of AM processes. Hybrid Bayesian Net- Criteria considered Targeted Value
works amalgamate empirical data with expert knowledge, therefore,
Industrial use of the ML Manufacturing process
augmenting dependability in decision-making, especially within intri- deployment
cate AM workflows. The specific methodology for Additive manufacturing
the investigation
Purpose of the study AM lifecycle, design, process planning, build, post-
2. Methodology
processing, quality, test and validation.
Keywords used Machine learning, 3D printing, Additive
This article used a systematic review as the research methodology. A manufacturing, Rapid prototyping, Artificial
systematic review is utilized to create a relevant and beneficial knowl- Intelligence, Deep learning, Design, Optimization,
edge base. This assessment considers previous research in the areas of Quality, Defect
Date of Publication Jan 2014 to May 2024
interest, as well as prospective topics for future inquiry. To address a

5
S. Inayathullah and R. Buddala Results in Engineering 25 (2025) 103676

Fig. 4. PRISMA flow chart-Systematic Review of ML Applications in AM: Publication Trends, Key Findings, and Future Directions.

Fig. 5. Number of Scopus publications connected to ML papers in the AM field from 2014 to 2024.

articles in the AM domain from 2014 to 2024 from various countries and kerf characteristics in the CO2 laser cutting process on polymeric ma-
territories. The United States has the most influence, followed by India, terials, which improves accuracy and increases productivity during
the United Kingdom, China, and Italy. cutting [26]. Recently, Sanchez et al. (2021) used ML to enhance pre-
diction of the thermohydraulic performance of shell and tube heat ex-
3. Findings changers that in turn demonstrated its potential for thermal systems
optimization [27]. ML was suggested by Elsheikh (2023) for its contri-
3.1. Applications of ML algorithms in different fields bution in predicting joint properties and tool failure detection during the
real process of friction stir welding applications [28]. Furthermore,
The demand for more complex parts has further led to massive Elsheikh et al. (2022) investigated low-cost bilayered structures for the
strides in fabrication techniques, with AM seeing a particularly strong improved performance of a solar still with an analysis on both using
surge. One of the main drivers in this evolution is ML, with its incredible APP/Costs, ML-based Predictions: on Water Yield. Shanmugan et al.
capabilities across many engineering fields (Fig. 9). For example, (2022). developed solar field distillation system [29]. Additionally, Devi
Alhawsawi et al. (2023) developed an ML algorithm for the prediction of et al. (2023) employed ML methodologies to predict moisture, calliper,

6
S. Inayathullah and R. Buddala Results in Engineering 25 (2025) 103676

Fig. 6. Various authors contributed to Scopus publications related to ML studies in the AM sector between 2014 and 2024.

Fig. 7. Subject categories with Scopus publications related to AM ML articles from 2014 to 2024.

and grammage for real-time paper quality assessment. The k-Nearest Besides, Abd Elaziz et al. (2024) predicted the freshwater yield of a
Neighbour approach has enhanced speed and precision in forecasting particular membrane desalination system using ML [32].
steam pressure performance and input variables [30]. Furthermore, The versatility of ML is further evidenced as shown by Mudhsh et al.
Devi et al. (2023) proposed some other models, which employ LSTM (2023) Achievement: Modeling the thermo-hydraulic behavior of helical
modeling with state-of-the-art algorithms for high throughput [31]. heat exchangers using ML and optimization techniques [33]. Djouider

7
S. Inayathullah and R. Buddala Results in Engineering 25 (2025) 103676

Fig. 8. Number of Scopus publications related to AM ML articles from 2014 to 2024 from various countries and territories.

Fig. 9. Applications of ML Algorithms in different fields.

et al. (2023) Using LSTM to study the dynamic response of Friction Stir causing maximum downtime [37]. Finally, Harmayanti et al. (2024)
processed nanocomposites with a case application on advanced material used Random Forest regressors applied to some cognitive manufacturing
characterization alongside ML [34]. Rajhi et al. (2023) done the problems, proved that ML could impose an opening for human computer
ML-based milling surface features in carbon fiber composites [35]. interaction optimization [38]. From this perspective, these examples
Elsheikh et al. (2022) presented the innovation of ML sees a new para- definitely exhibit the kind of change facilitated by ML and hence it es-
digm shift in cost and performance estimation for bilayered solar stills tablishes a compulsion supporting the incorporation of ML into AM
which has resulted in more catalytic changes encouraged by these processes creating improvement and productivity.
techniques that took place around several fields of engineering [29].
Muqeet et al. (2023) offers a Q-learning-optimized PID control system 3.2. Integration of ML in AM
that utilizes piezoelectric feedback to improve surface quality in turning
operations. The technology adaptively modifies cutting depth, guaran- A thorough analysis of the current state of ML applications in
teeing resilient, steady performance and enhanced outcomes across numerous AM domains is presented. ML in the DfAM may be used to
diverse materials [36]. Elminir et al. (2024) introduce an LSTM-AE create optimal topological designs and new high-performance meta-
model for forecasting CNC tool wear utilizing sophisticated character- materials. Modern ML techniques may be used in AM processing to
istics such as entropy and interquartile range (IQR). The method, trained optimize process parameters, investigate powder dispersal, and detect
on the PHM10 dataset, accurately predicts wear and Remaining Useful in-process defects. The possibility of AI and ML enhancing the 3D
Life (RUL). It allows for scheduled repairs to be done efficiently without printing process is important because both have a high potential to

8
S. Inayathullah and R. Buddala Results in Engineering 25 (2025) 103676

improve design optimization, defect detection or even just bringing it the strategies that may be used to improve the efficiency of modern
closer control. ML algorithms can then take this data from massive da- manufacturing. First, we’ll talk about the advances in DL and the ad-
tabases of previous builds and use it to help optimize design parameters vantages it has over traditional ML. Computational methods that are
in the current build process-resulting in less material waste and better based on DL are subsequently introduced with the specific objective of
performing builds. In the manufacture of AM, ML may help with pre- enhancing the performance of manufacturing systems. Our analysis in-
manufacturing scheduling as well as product quality evaluation and volves the comparison and contrast of numerous representative DL
assessment [39–42]. AM is increasingly concerned about data security models [44]. Research in this field has produced an algorithm that uti-
due to the potential for data breaches that could be caused by ML lises big data analytics to anticipate production costs for current projects
techniques. AI-based systems are capable of monitoring in real-time based on previous costs of comparable operations. This pricing analytics
during print and can detect defects immediately which means that service may help makers simplify their task bidding process [45].
product quality and consistency is greatly improved. There are also ML
methods, specifically designed to improve predictive maintenance by 3.3. Key mechanical properties in AM via ML integration
examining the data of machines in order to predict a failure and prevent
downtime (predictive failures) [43]. Table 3 shows the basic mechanical properties in AM, tensile, elas-
One interesting advantage of integrating ML with AM is the ticity, hardness and fatigue resistance and the benefits that ML generates
increased process optimization that allows correlating input variables, i. in their evaluation to optimization. Using ML, these models predict
e., temperature and speed at which each new layer is laid down while mechanical responses to be expected by analyzing printing parameters
helping to improve printing parameters in order to generate prints more as well as material behaviors for part quality and reliability. Some of
quickly or reduce material waste. This helps in quality control, as the them predict tensile strength through neural networks, such as many
system can predict possible defects well before onset and allow you to others consistently need to be controlled in adapting control mode like
maintain proper product consistency by analyzing sensor data on a density and thermal stability. Such an integration of ML enables process
timely basis. It also enables the improvement of predictive maintenance, enhancements, real-time control and post-processing as well, which
as ML can determine what historical data to analyze and how long it leads to improving the performance and automation in AM systems.
changes from normal so that downtimes are minimized in cost and time.
An automated design evaluation reduces (some) difficulties by making 3.4. Applications of ML with AM
predictions on the designs produced before manufacturing begins,
whereas ML allows for faster material development by predicting ma- ML and AI have had a significant impact on AM. We are transitioning
terial properties based upon compositional data. Every time a part is to their new manufacturing procedures at various stages. These cutting-
printed, the massive datasets produced throughout an AM process are edge approaches aid in the improvement of processing parameters,
distilled into actionable real-time information to enhance part design defect detection, part property prediction, quality control, shape dif-
and production. ML also allows greater personalization for design ference management, quality forecasting, closed-loop control, and cost
optimization to suit certain user requirements, and at the same time estimation (Fig. 10) [53]. We may utilise 3D printing, also known as AM,
results in cost control. Finally, it increases simulation accuracy to better to create complex, one-of-a-kind components which are nearly unfea-
predict real-world thermal and stress performance of a part by keeping sible to manufacture using conventional material removal methods.
the model up-to-date with field data. Making things has gotten more difficult as our material possibilities and
This study provides an extensive investigation of ML algorithms and creative freedom have expanded. ML and AI can help address these

Table 3
Essential mechanical properties and the importance of ML in AM [46–52].
Mechanical Description Role in AM Significance of ML ML Example
Property

Tensile Strength The maximum stress that a material Guaranties longevity and Analyzes process parameters, including Neural networks forecast tensile
can endure before it breaks. mitigates mechanical laser power and layer thickness, to strength based on historical print data.
malfunction. forecast the tensile strength.
Elastic Modulus Assesses material rigidity or resistance Guarantees structural Enhances printing settings to provide SVMs forecast elasticity according to
(Young’s to elastic deformation. integrity, particularly in the requisite rigidity and mechanical material and process parameters.
Modulus) load-bearing components. stability.
Fatigue Resistance Capacity to endure several stress cycles Essential for dynamic Forecasts fatigue resistance based on The RFs model predicts fatigue
without failing. applications such as stress data and layer quality. resistance utilizing previous stress cycle
prosthetics and aircraft. data.
Hardness Resistance to superficial distortion or Guarantees longevity in Identifies imperfections that diminish CNNs identify micro-voids in the
abrasion. high-friction components hardness during layer deposition. printing process that influence
such as gears. hardness.
Impact Strength Capacity to absorb energy without Crucial for automotive and Enhances characteristics to augment RL enhances impact strength through
(Toughness) sustaining fractures under abrupt military components. toughness through the modulation of process modifications.
stresses. energy inputs.
Density and Density is defined as mass per unit Influences mechanical ML-based predictive control maintains Predictive models regulate laser power
Porosity volume, whereas porosity pertains to strength and fatigue uniform density by overseeing printing to provide consistent part density.
the spaces present inside a substance. resistance. parameters.
Thermal Comprises thermal conductivity, Essential for high- Forecasts the response of components to Physics-informed ML techniques
Properties thermal resistance, and expansion. temperature applications thermal and mechanical stress via simulate thermal dynamics and forecast
such as aerospace. simulation-based models. conductivity.
Post-Processing Improves mechanical characteristics Improves strength and ML algorithms suggest ideal post- Bayesian optimization refines
Optimization by heat treatment or surface finishing. durability for industrial use. processing parameters to improve annealing parameters for alloy
attributes like toughness and fatigue components.
resistance.
Real-Time Process Guarantees process uniformity Ensures uniformity in part Predictive ML dynamically modifies Adaptive control systems adjust laser
Control throughout production. quality and mechanical printing parameters to prevent faults power and speed during the printing
properties. and uphold quality. process to guarantee consistent
strength.

9
S. Inayathullah and R. Buddala Results in Engineering 25 (2025) 103676

Fig. 10. Applications of ML approaches in AM.

difficulties. They help to automate and enhance numerous phases in the in ML Methods for Design Assessment.
AM process. ML is often used in AM to improve processing parameters. Yao et al. (2017) proposed a mixed ML strategy in which hierarchical
Scientists used ML techniques to build models. Some examples of ideal clustering is used to code AM aesthetics and specific components,
processing parameters predicted by these models are the speed of the resulting in the creation of a dendrogram. A supervised classifier is
laser power scanning and layer depth. This allows us to achieve the trained on samples from current industrial applications to determine the
necessary component attributes while minimizing faults. best sub-cluster in the dendrogram comprising suggested AM design
ML significantly influences the identification and resolution of flaws envisions. This technique, on the other hand, only considers
in 3D printing. Fig. 8 illustrate the utilizations of ML methodologies in functionality-centric design knowledge, such as objectives, loadings,
AM. Scientists have created computer vision systems and ML tools to and characteristics, when making AM design feature suggestions [58].
check for problems. This lets them watch and spot issues while things are Williams et al. (2019) investigated how ML can automate routine or
being made. It makes the end product better and saves time and money time-consuming engineering duties when abundance of data is avail-
compared to checking by hand. People also use ML to guess how parts able. Their study uses "artificial" design repositories to test how varia-
will turn out and if they’ll have the right shape. This helps makers deal tions in dataset properties affect the accuracy and generalizability of
with problems before they happen. Also, combining ML with a product’s neural networks. They employ a 3D convolutional neural network to
whole life cycle can help determine how 3D printing affects the envi- predict reckonable production parameters from voxel-based part ge-
ronment. This speeds up the design process and makes the technology ometry, with an emphasis on AM material extrusion. The study found
more environmentally friendly. As 3D printing advances, combining ML that when compared to a baseline multiple linear regression model, this
and AI will become increasingly vital. It will assist to improve the pro- neural network performed better, particularly for orientation-
cess, maintain good quality, and uncover cost-cutting opportunities independent indicators, and less standardised datasets as a whole. Ac-
[54–56]. Table 2. lists the various ML approaches used in the AM cording to all construction metrics, the CNN outperforms the baseline
process. linear regression model [59].
Guo et al. (2020) introduce a semi-supervised graph neural network
(GNN) approach for constructing architected material topologies,
3.5. Design assessment in the workflow of AM introducing a novel AI-driven paradigm for material design and opti-
mization. Traditional techniques address boundary value problems
This review focuses on transformative contributions that ML has (BVPs) under load and displacement constraints. GNNs anticipate 99 %
made to material optimization, topology prediction, and defect mitiga- of node loads with just 1 % of load levels. This technique, along with a
tion in AM design workflows.ML is employed to optimize material truss thickness algorithm, allows for multiscale material design under a
design and topology in AM. Some design processes have been enhanced variety of situations. This novel method improves designs with little
using ML models that combine Design for Additive Manufacturing sensor data under complicated loads, demonstrating significant poten-
(DfAM) concepts while minimising material waste. They may assess tial in AM, architecture, and civil infrastructure [60].
geometry correctness, material qualities, and any faults, allowing for Omairi et al. (2021) emphasize the need to recognize AM with
adjustments to be made at any time. As a result, this relationship im- minimal material waste as a paradigm shift towards evolutionary pro-
proves product quality while also lowering costs, resulting in new de- gression in the manufacturing industry. The authors reviews the heu-
signs with complicated geometries, therefore supporting sustainable ristic algorithms and predictive models that aim at making AM
manufacturing [57]. Table 4 shows a summary of the literature covered intelligent, efficient, and sustainable. It identifies current trends, future
opportunities, gaps, and requirements in AM in terms of inter-industry
Table 4 cooperation, training, education, and technology transfer aligned with
Summary of ML Methods for Design Assessment. Industry 4.0. Their review specifically looks at different ML algorithms
S.No AM application ML technique used Reference for finding flaws in AM cyber-physical systems and suggests areas for
1 Design feature SVM [58] further research, such as making the systems easier to print, lowering the
2 Design feature ANN [59] complexity of the designs, speeding up the prefabrication process,
3 Topology optimization GNN [60] improving real-time control, making the systems safer, and finding flaws
4 Design feature RF, DT, ET [61] more quickly [61].
5 Design feature CNN, SSM, PDM [62]
6 Design feature GANs, ANNs [63]
Wu et al. (2022) argue that one form of continuously operated wet
7 Design feature Five-fold cross-validation scheme [64] hydrodynamic technology can prepare chitosan-based nanoparticles
8 Design feature DDIM, GANs [65] (CNPs). They find application in drug delivery, skin care, and food
9 Design feature CAEs, TL and GAN [66] preparation. The method employed two tools: a 3D-printed multi-inlet

10
S. Inayathullah and R. Buddala Results in Engineering 25 (2025) 103676

vortex mixer (MIVM) and a peristaltic pump for flash nanoprecipitation also enhancing biological activities including angiogenesis, anti-
(FNP). It is not incorrect to assert that the high polymer dispersity index inflammation and ROS scavenging, which helps to speed up wound
(PDI) and large nanoparticlphohae size of CNPs are crucial for their healing. This bioinspired method represents a viable platform for ther-
application, as well as for ensuring their stability and clarity. The results apeutic use in acute and chronic wound healing. The design of DNA-
revealed that RF’s most accurate model is 91 %. The MIVM instrument induced biomineralization-based bioinspired 3D-printed hydrogels is
can provide the desired production conditions, such as high-volume displayed in Fig. 12 [64].
manufacturing of nanoparticles at the nanoscale structural level, Yangue et al. (2024) show a new development in evolving diffusion
ensuring their quality, and representing an apt and green approach to generative models: the denoising diffusion implicit model (DDIM). This
nanoparticle manufacture [62]. model is used for layer-by-layer picture enhancement and supervising in
According to Burge et al. (2022), the use of X-rays as a guiding tool is 3D printing. To improve layer-wise AM image augmentation, the sug-
seen as a significant restrictive factor in the process of regular knee re- gested models include a pair kernel based distance measurement tech-
placements implant selection since the surgeon can only choose limited niques into the DDIM architecture. Among the suggested measures are
models and sizes, which sometimes do not correspond to ideal bones,
causing fitting problems and, therefore, consequent revisions. Specif-
ically, the use of personalized implants played a major role in improving
the whole process but required extensive preoperative exams, longer
lead times, and higher costs. Consequently, various limitations existed
that protracted their use. Burge et al. (2022) proposed using CNN and
statistical shape models to customize mass-produced implants by
generating calibrated X-ray images. The models are created by the AM
method and this greatly improves the suitability and thus, lowers the
number of revisions. This streamlined process attained a femur mea-
surement of 1.00 mm and a tibial measurement of 1.07 mm for 30
participants [63].
Zheng et al. (2023) demonstrate progress in using AI to construct
aeroplane wings using a 2D dragonfly wing skeleton. The cellular core of
aviation wings was developed using ML and examined using reciprocal
diagrams. Fig. 11 shows the graphical statics, ML, and applying drag-
onfly wings to aeroplane wing construction. This strategy resulted in the
development of lightweight, high-performance wing structures. The
wing of a dragonfly possesses an exceptional and precisely defined dual/
reciprocal diagram, suggesting that its internal network forms a deter-
minate structural framework. The structural elements in the fabricated
wing design closely match the dimensions of those found in the natural
dragonfly wing. The wing designs inspired by dragonflies, generated
using the proposed methodology, exhibit up to a 25 % increase in out-of-
plane stiffness compared to conventional wing designs [63].
Kim et al. (2023) presents a novel method for 3D printed bioactive
hydrogel coverings for diabetic injury healing that employs marine-
derived ingredients. DNA-induced biosilica and Salmon sperm DNA
are integrated into 3D-printed alginate hydrogel dressings, resulting in
porosity for exudate absorption and mechanical tunability for form Fig. 12. Schematic of DNA-induced biomineralization-based bioinspired 3D-
consistency. These dressings protect wounds from deformations while printed hydrogels [64].

Fig. 11. Graphical statics, ML, and applying dragonfly wings to aeroplane wing construction [63].

11
S. Inayathullah and R. Buddala Results in Engineering 25 (2025) 103676

KID-IS, which integrates m-KID with the inception score (IS), and m-KID, the optimal process parameters (powder feeding rate, scanning speed,
a variant of the kernel inception distance. These new combinations have laser power) necessary to meet specific geometrical criteria. Fig. 13
a lot of potential for keeping things the same and similar in AM layer- depicts the average accumulation height (h), breadth (w), and depth (d)
wise picture enhancement while also looking into any changes that associated with each of the 30 investigational circumstances. The blue
weren’t seen. This case study examines six examples of FFF using metal marks denote actually observed averages, whereas the orange markings
and polymer materials. The findings designate that the DDIM/KID-IS denote the estimates made through the ANN. The ANN accurately
model and DDIM/m-KID model outperformed the four standard tech- computed these parameters, yielding mean absolute proportion errors of
niques. The mentioned models include the widely recognized denoising 5.5 % for powder feeding rate, 5.8 % for scanning speed, and 2.0 % for
diffusion probabilistic models (DDPMs) as well as three additional GANs laser power [68].
[65]. Mohammadi et al. (2021) evaluated the performance of various ML
Mohseni et al. (2024) provide a procedure to enhance the manu- methodologies for in situ defect detection in the laser powder bed fusion
facturability of topology-optimized (TO) structures in AM. Editing of (L-PBF) process with the goal of improving product quality and process
geometric constraints and classification of building orientations are reliability. The researchers used an acoustic emission sensor to detect
introduced to mitigate stresses and reduce the necessity of support defects at a very low cost by capturing elastic wave generation through
structures. CAEs, TL, and GAN are demonstrated in the automation of printing. They fabricated H13 tool steel samples using three pre-
DfAM. FEA confirms 50 % less stress in the edited parts. These AI- calibrated defect sizes, capturing AE signals throughout the entire
predicted BOs achieved a high level of accuracy, making the imple- build process. This study’s major methodologies include hierarchical K-
mentation of SS in laser-based powder bed fusion (LB-PBF) samples means clustering, DL neural networks, PCA, Gaussian Mixture Model
unnecessary. These all result in providing an automatic platform for (GMM), and VAE. Supervised DL detected the quality trend; however,
DfAM that can facilitate efficient design and manufacturing of complex the VAE methodology revealed a new defect detection method that did
TO parts [66]. not need training the model with diverse materials [69].
Rezapour Sarabi et al. (2022) developed an AI framework to antici-
3.6. Defect detection in the AM process pate and analyse the characteristics of 3D-printed microneedles,
enabling the creation of novel medical systems and applications. FDM
ML algorithms improve flaw detection in AM by analysing data from 3D printing was utilised to create biodegradable microneedles, which
several sensors and monitoring systems. These approaches can detect were then chemically etched for increased precision. DL was utilised to
patterns and abnormalities that signal flaws, such as porosity, fractures, manage quality and detect anomalies in Microneedles. A data collection
or dimensional irregularities, in real time. ML enhances the accuracy of 10 microneedle designs and etching exposure dosages was utilised to
and speed of flaw detection by using methods like as decision trees, develop ML models to extract similarity metrics to forecast new fabri-
SVMs and neural networks resulting in higher quality and more cation results when parameters were modified [70].
dependable AM processes. Table 5 presents a summary of the literature Jyeniskhan et al. (2023) developed a digital twin system for AM that
studied on ML approaches used for detecting flaws. uses an ML method (EfficientDet-Lite object detection model) for flaw
Wu et al. (2016) proposed a method for identifying malicious infill detection and a Unity client user interface. This system provides real-
failings in 3-D printing processes. It explores these flaws, extracts time feedback to operators, allowing them to make alterations before
characteristics from simulated pictures, and runs a test with both non- fabricating defective goods. The OctoPrint interface is used by the 3D
defect infill photos and defect training images. The data is then used printer to collect data, such as video and images, and to intentionally
to train a pair of ML algorithms, the Naive Bayes Classifier and the J48 make the printed items flawed in order to teach the ML model [71].
DTs. The findings reveal that when it comes to classification, J48 DT Chowdhury et al. (2023) provides a data-driven strategy for identi-
achieve an accuracy of 95.51 % and the Naive Bayes Classifier a mere fying 3D printer problems, which involves gathering information on
85.26 % [67]. healthy, bed failure, and arm failure scenarios. The RF and XGBoost
Caiazzo et al. (2018) conducted research on the laser direct metal ensemble model outperformed the individual models with an accuracy
deposition (DMD) technique to deposit 2024 Al alloy onto T3 temper of 99.75 %. Using the Shapley additive explanations library boosts the AI
plates that were 5 mm thick. The primary goal was to control the geo- model’s trustworthiness [72]. Metal AM items can have internal faults
metric characteristics of the metal trace that was deposited. The study while achieving complicated shapes and decreasing costs.
employed an artificial neural network (ANN) ML technique to ascertain Non-Destructive Testing (NDT), notably CT, is expensive and
time-consuming for defect analysis. Bordekar et al. (2023) report an XAI
system for automated flaw discovery and characterisation. The
Table 5
Overview of ML methods used for detecting defects.
approach was evaluated on Ti6Al4 V specimens using SVM with an AUC
of 0.94 to categorize CT image pixels as pores or inclusions, followed by
S. AM application ML technique used Reference
DBSCAN clustering and convex hull characterisation. XAI’s usefulness in
No
NDT and AM component assessment was shown by its automated flaw
1 Infill flaws detection in Naive Bayes Classifier and J48 DT [67] identification [73].
AM
2 Defect detection and ANN [68]
Demetgul et al. (2024) have come up with a new automated ML
quality control (AutoML) method for detecting linear phases in the machine tools and
3 Defect detection using Hierarchical K-means clustering [69] AM equipment, especially semisensors. AutoML uses techniques of
AE sensor supervised DL neural network, electric motor current data to diagnose misalignment problems with
PCA-GMM and VAE
high accuracy. The use of ensemble learning methods improves the
4 Anomaly detection Classical ML method [70]
5 Defect detection Efficient Det-Lite object detection [71] classification accuracy whereas optimizing the hyperparameters en-
model hances it even more. In the future, it is imperative to monitor motor
6 Defect detection RF and XGBoost [72] current signals and to use AI technology for the operation of machinery
7 Defect detection and SVM, EXplainable AI (XAI) [73] and its diagnostics, notes the research [74].
characterization
8 Defect detection and SVM, GB, and AUTOMLP [74]
ML is widely utilized in machining to forecast tool wear, surface
Production control roughness, and other crucial features. Danish et al. (2024) forecast flank
9 Defect detection LR, SVM, RF, MLP, [53] wear (Vb) to assess tool wear and class separation. A prediction model is
10 Design development ANN [75] given for MQL, cryogenic, and MoS2+MQL situations.ML approaches
and defect detection
like support vector machines, linear regression, multilayer perceptron

12
S. Inayathullah and R. Buddala Results in Engineering 25 (2025) 103676

Fig. 13. Investigational average (blue markings) and ANN estimated (orange markings), a) deposit breadth (w), b) altitude (h) and c) depth (d) for every one of the
thirty testing scenarios [68].

and RF, provide predictions. The research found that the multi-layer
Table 6
perceptron (MLP) outperforms conventional classification approaches,
Outline of ML methods employed in process control and optimization of the AM
achieving over 95 % prediction accuracy across testing and training
process.
datasets. Despite the lack of adequate information, the proposed
approach is capable of recognizing circumstances and providing rec- S. AM application ML techniques used Reference
No
ommendations for machining strategies [53].
Ghimire et al. (2024) in their publication illustrate the way multi- 1 Process control MPC [76]
2 Process control CNN(AlexNet) [77]
functional composites can perform multiple functions successfully.
3 Parameter optimization Conventional autoencoder, [78]
However, the companies find 3D-printed multifunctional carbon com- Clustering algorithm, RF
posites for airplane requirements to be difficult due to substances and 4 Build time, speed prediction ANN [79]
process characteristics, anisotropy, and absence of standardization. AM and Quality print
and composites are neither advanced to the extent of be safety-critical 5 Process control ANN & ANN-GA [46]
6 Optimization of print LR, DT, SVR, RFs, GRB, and [80]
nor are their drawbacks confidently revised. Ghimire et al. (2024) sug- quality XGB
gest the application of the ML approach to emergent such composite 7 Optimization of print DT, RF and DL [81]
material, which could be employed for airplanes, unmanned aerial quality
systems, and spacecraft with the help of AM technology beside ML. To 8 Process control RL [82]
9 Production and process Self-organizing map (SOM) [83]
this end, they give a detailed explanation of an ANN, that is part of an
control algorithm
AM strategy. A powerful tool such as ANN architecture combined with 10 Production and process RF [84]
an added in-situ 3D printing method proposed by the authors was used. control
This tool provides an innovative way to overcome the existing limita-
tions related to multifunctional materials’ properties, production
methods, and failure modes by comparing AI ANN utilizing anticipa- and validates the prognostic potential of ML models derived from pro-
tions and historical data [75]. While ML models achieve high accuracy, cess data. The findings indicate that state-space models may forecast
these models often struggle to adapt to different datasets and are many process trends and product quality attributes across long time horizons,
times not suitable for real-time deployment in an industrial setting. allowing MPC to function optimally [76].
However, these limitations motivate the development of standardized Khanzadeh et al. (2018) address the issue of geometric accuracy in
data collection, robust training methodologies, and scalable solutions to AM, specifically for complicated shapes. They point out that current
fit within an industrial AM environment. geometric dimensioning and tolerancing (GD&T) methods are insuffi-
cient for such components. The work uses a self-organising map (SOM),
an unsupervised ML technique to evaluate geometric discrepancies in
3.7. Process and performance optimization in AM laser-scanned data. The notion is that SOM clusters represent distinct
geometric aberrations associated with AM process circumstances. The
ML algorithms are critical for improving the process and perfor- technique, which was tested on FFF components, effectively visualised
mance of AM. Manufacturers may use these algorithms to fine-tune and quantified the link between process parameters and geometric
settings, improving print quality while maintaining quality control. precision while lowering the number of data points necessary to <3 % of
This optimization guarantees that the prints are consistent and high- over 1 million [83].
quality. Furthermore, ML can forecast speed and estimate construction Kappes et al. (2018) printed about 3600 samples to characterise
time, resulting in more efficient production plans. Ultimately, applying Inconel 718 construction characteristics using laser powder bed fusion
ML into AM greatly improves print quality and operating efficiency. (L-PBF). They established a relationship between pore creation, posi-
Table 6 summarizes the literature examined on ML algorithms used for tion, part orientation, and utilisation of reused powder. This data aided
AM process control and performance optimization. the creation of an RF Network ML model for process-property and
Stoyanov et al. (2017) examine the use of predictive models for process-structure interactions. The work examined how procedural
process monitoring in AM technologies, including 3-D inkjet printing, to stages in l-PBF impact porosity, namely lack of fusion (LOF) and keyhole
guarantee the quality of electrical devices. This approach replaces flaws, and how the potential for data gathering and ML might uncover
typical post-production inspections. The authors propose a methodology subtle correlations between input variables and output parameters [84].
that combines a model-based approach with ML algorithm to effectively Gardner et al. (2019) conducted ML optimizes parameter settings for
manage product quality throughout manufacturing and achieve model quality 3D-printed parts. The printed parts of outside walls were pho-
predictive process control (MPC). This study stands out because it spe- tographed with a Nikon D3200 DSLR photographic camera and an 85
cifically concentrates on using data-driven prognostics, which involves mm macro lens and a dark background to obtain data. In MATLAB
using state-space modeling to analyze the dynamic behavior of 3D inkjet r2018a, AlexNet was utilized for monitoring. Global parameters at 40
printing for electronics manufacture. The principal application is the 3D mm/s created components with 14 % more faults than quality-optimized
printing of conductive lines for electrical circuits, which demonstrates

13
S. Inayathullah and R. Buddala Results in Engineering 25 (2025) 103676

prints. Quality-optimized components printed 30 % faster and had 13 % whilst 3D printing allows for the rapid fabrication of complex structures.
higher quality than global settings, showing that operator priorities Nonetheless, optimizing printing settings remains a hurdle, resulting in
determine the quality-speed trade-off. Fig. 14. compares the perfor- delays and resource losses. Rahmani Dabbagh et al. (2022) improve
mance of the tool (quality optimised, print time, and general settings) parameters by combining ML with 3D printing and employing a user-
for different parameter values [77]. friendly graphical user interface. Fig. 16 shows a completely inte-
Silbernagel et al. (2020) illustrate the utilization of ML as a substitute grated deep neural network designed for metal AM. The study employed
approach for determining this specific set of process variables. This nine CAD images and a self-created difficulty score to compare the
experiment primarily investigates the laser powder bed fusion printing printed outputs against CAD drawings labeled in four different ways.
process of solely copper. Visual images were gathered throughout the Training eight ML algorithms on 224 data points revealed that the
conventional process of parameter optimization. The photos were "gradient boosting regression" model performed best, with an R² value of
divided into segments and inputted into a convolutional autoencoder. 0.954. Their ML-embedded graphical user interface (GUI) allows clients
The resulting clusters were then analyzed to identify the clusters that to enter designs and desired printing parameters to anticipate print
most accurately indicated a superior outcome. The clusters were similarity, minimizing the necessity for trial-and-error, speeding up
manually assessed for quality, and the outcomes were subsequently production, and avoiding wastage [80].
compared to the distinctive set of parameters. The machine-learned Chen et al. (2023) employed ML algorithms to estimate the efficacy
grouping and subsequent counting accurately replicated the findings of different biomaterial inks used in 3D printing, with the objective of
made during the conventional parameter optimization method [78]. eliminating trial-and-error testing and speeding up research. They
Oehlmann et al. (2021) conducted a study used ANN to accurately analyzed 210 ink formulations and determined that the method was
model FFF processes, helping to understand the key factors that impact effective. While RF was the most accurate in predicting printability, DL
print quality. The research used ANN models to successfully predict the provided a more sophisticated understanding. Printability, defined as
structural strength and failure points of printed parts, making testing precise control over ink deposition and form retention during 3D
easier and improving overall print quality. Fig. 15 shows the data from printing, has a direct impact on the end product’s quality and func-
filament velocity and force sensors during cylinder print. When properly tioning. This investigation demonstrates the potential of ML to expedite
trained, the ANN models mimicked the functions of a force sensor unit in the creation of biomaterial inks, notably in biomedical applications
an FFF device, providing technological innovations readily available to a (Fig. 17). This paves the path for the rapid development of tailored
wider audience. By analyzing real-time printing device information, the bioinks, particularly for 3D printing applications [81].
ANN algorithm accurately predicted nozzle behavior, facilitating the Mattera, et al. (2024) addresses the issues in optimal control design
identification of the optimal settings for process speed and print quality for industrial processes caused by stochasticity, nonlinear dynamics, and
[79]. restrictions. The authors propose using RL to regulate multi-input, multi-
Meiabadi et al. (2021) researched effective and most economical output non-linearly restricted systems. While RL is widely used in ro-
ways of 3D printing the tensile test samples in the desired thickness botics, its applications in manufacturing have been restricted. They
using FFF. The temperatures of the extruder, infill percentage, and layer suggest a procedure-based reward function (a software solution) and
thickness were correctly chosen by the use of the response surface create their own lower order model of the surrogate simulator on WAAM
method (RSM). The manufacturing cost, component thickness, and to account for nonlinear couplings. Although real-time implementation
toughness were the three parameters which were predicted through is difficult, the study demonstrates that RL holds considerable promise in
ANN and the ANN-GA. It was established that the combination of the manufacturing. Tyberg stated that future study would include real-
hybrid ANN method with GA enabled a 7.5 % improvement in durability world applications and comparisons with different control systems
accuracy, an 11.5 % increase in component thickness and a 4.5 % [82]. RL techniques for non-linear control in AM are shown to be
reduction in manufacturing cost. The PLA produced the cheapest and promising ways to overcome constraints of traditional optimization
most convenient specimen since it was properly optimized [46]. models. Enabling more efficient, more precise, more adaptive
ML and 3D printing are growing disciplines with rapid growth. ML manufacturing processes, these advancements lay the ground for
allows computers to gain knowledge and improve performance by broader industrial adoption.
evaluating data without requiring considerable human intervention,
3.8. Quality control in AM

ML algorithms improve quality control in AM by optimizing pro-

duction processes and lowering costs. In cyber manufacturing, these
algorithms analyze large datasets to discover and forecast faults,
resulting in consistent performance. ML maintains high quality stan-
dards by continually monitoring and modifying parameters, resulting in
consistent and efficient manufacturing results. Integrating ML into AM
enables greater quality control, leading to high-performing products and
more efficient processes. Table 7 is a summary of the literature covered
in ML Methods for Quality Control in AM.
Wasmer et al. (2019) discuss the shortcomings of existing quality
monitoring systems in AM, which rely on temperature readings and
high-resolution images but fail to capture underlying physical processes.
They offer a novel approach that combines acoustic emission with RL to
check quality in real time. Acoustic emission records subsurface dy-
namics during the AM process, and RL evaluates this data to determine
quality. This combination technique was tested on a commercial AM
machine, resulting in stainless steel 316 L cuboid geometries with varied
porosity levels. The approach displayed good classification accuracy,
distinguishing distinct acoustic signals corresponding to different qual-
Fig. 14. Global parameter settings versus tool performance in optimizing print ity levels, demonstrating its potential for dependable, real-time AM
quality, time, and quality [77]. quality monitoring [85].

14
S. Inayathullah and R. Buddala Results in Engineering 25 (2025) 103676

Fig. 15. Data from filament force and velocity sensors during cylinder printing [79].

manufacture. On-demand ODF manufacturing and quality control were

accomplished with the help of 3D printing and ML. DIW can create
complicated ODF structures with thicknesses under 100 micrometers.
ML algorithms were used to categorize ODFs based on their primary
component utilizing near-infrared (NIR) spectra. A linear discriminant
analysis model classified ODFs 100 % accurately. Using partial least
squares, paracetamol, caffeine, and theophylline ODFs exhibited dosage
verification coefficients of 0.96, 0.99, and 0.98. 3D printing, NIR, and
ML speed ODF creation and verification. Automatic in vitro testing using
machine vision shows the possibility for a completely automated ODF
procedure [87].
Nasiri et al. (2021) investigated ways to enhance FFFquality in pol-
ylactic acid to boost toughness and reduce production costs for 3D-
printed tensile test samples. They employed the response surface
method (RSM) to regulate the temperature of the extruder, choose the
proportion of infill, and specify the thickness of each layer. They
developed ANN and ANN-GA models to calculate toughness, component
thickness, and manufacturing costs. It was shown that ANN-GA out-
performed ANN in terms of toughness (7.5 % improvement), component
Fig. 16. A completely linked deep neural network for metals AM [80]. thickness (11.5 %), and manufacturing cost (4.5 %) modeling accuracy.
The enhanced specimen was both affordable and bendable, enhancing
Kunkel et al. (2020) developed a standardized, efficient, and cost- the popularity of printed PLA goods [88].
effective quality control technique for metal powder bed fusion AI in digital manufacturing models can enhance efficiency and
(PBFM) using process monitoring data. Deep CNNs have the potential to quality, especially in metal AM. DTs can independently regulate and
reduce QC costs and throughput times by recognizing patterns that optimize physical processes using AI. High-quality digital models and
humans cannot see. ML was used to enhance pattern identification in fast-solving surrogates make DTs effective. Gunasegaram et al. (2021)
melt pool photos, resulting in 98.9 % accuracy in identifying 9280 un- emphasizes DTs in AM and high-fidelity multiscale-multiphysics models.
seen layer photographs. This provides an automatic quality control It lists technological issues like integrating models across scales and the
strategy for PBFM [86]. absence of experimental data, as well as non-technical issues like stan-
O’reilly et al. (2021) created Orodispersible films (ODFs), which dardization and collaboration. Model quality and progress are improved
have therapeutic and economic benefits but need tailored, sustainable with solutions and a shared development plan [93].

Fig. 17. ML Algorithms: DT, RF, and DL for Hydrogel Printability [81].

15
S. Inayathullah and R. Buddala Results in Engineering 25 (2025) 103676

Table 7 the world’s fifth most prevalent cancer, is the third leading cause of
Summary of ML Methods for quality control in AM. cancer-related mortality. The processes of bioprinting, 3D printing, AI,
S. AM application ML technique used Reference along with machine intelligence are the ones that are emerging right
No now, and that will help to deal with the management of HCC. Discourse
1 Quality RL [85] today is that these technologies are also being dealt with by the "chil-
monitoring dren" of the technology’s formers. The utilization of AI with 3D printing
2 Quality control CNN [86] is the standard and unquestionable but overlooked method of diagnosis
3 Production and PCA [87] and treatment. The doctors are the real actors here, so, by the way, are
quality control
4 Production and ANN-GA [88]
you referring to them before finding out more details about these
quality control technologies which ultimately help in the treatment of HCC [94].
5 Production and ANN, SVM and RFs [89] Khusheef et al. (2023) use DL signal-to-image encoding for AM
quality control process monitoring utilizing multi-sensor data fusion (MSDF). Tradi-
6 Production and ANN, KNN, and SVM [51]
tional Kalman filtering and AI-powered solutions are discussed. The
quality control
7 Quality control MSDF [90] work uses recurrence plot encoders, Markov transition fields, and Gra-
8 Production and SurpriseAF-BO, CTGAN [91] mian angular fields to develop decision fusion models, feature models,
quality control and data models. Using Delta 3D printer data, feature-level fusion using
9 Quality control HSI [92] a hybrid convolutional and recurrent model yields the maximum accu-
10 Quality control FastForestOva, [48]
LbfgsLogisticRegressionOva,
racy (up to 99.6 %) with resilience and stability, whereas data-level
SdcaMaximumEntropyMulti, fusion models take less time but have lower accuracy [90]. Although
LightGbmMulti, FastTreeOva the DL based approaches have the ability to leap in the AM defect
detection, they behave as black box systems and are not easily inter-
pretable, which makes them not ready for use in the critical industries.
Ong et al. (2022) overcome the difficulty of lengthy empirical
By adding transparency and insight to model decision through XAI
formulation procedures in 3D printing (3DP) for healthcare by using ML
methods, it builds trust and makes it possible to scale model imple-
to anticipate formulation results. The researchers utilized internal and
mentation. While this quality control gap likely exists for future XAI for
literature-derived data on HME (hot melt extrusion) and FDM (fused
AM, efforts in this area can work to help build a system that can achieve
deposition modeling) 3D printing to generate a well-balanced dataset
high accuracy and high interpretability.
consisting of 1594 compositions. Fig. 18 depicts the Radar map that
Raihan et al. (2024) have reshaped Metal Additive Manufacturing
displays the printability metrics findings. A total of 84 % accuracy was
(MAM) by permitting it to generate different materials, which would
achieved by the enhanced ML models in predicting printing capabilities
have otherwise not been possible, fast prototyping, and innovative de-
and filament mechanical characteristics. In terms of processing tem-
signs without waste. While enjoying these benefits, obviously, product
peratures, HME was predicted with a mean absolute error of 5.5 ◦ C and
quality comes first in manufacturing. It’s only adding AI into MAM that
FDM with an error of 8.4 ◦ C. These findings beat prior models that used
can solve this problem. The facilitators claim that the SurpriseAF-BO
smaller, more unbalanced datasets, emphasizing the need of structured
surprise-based sequential learning protocol predicts the relationship
and varied data for optimal ML performance [89].
between the process parameters and the melt pool features with the least
Gor et al. (2022) emphasize the importance of density in powder-
amount of data. Their way of melting the pools depth, breadth, and
based AM, particularly for stainless steel 316 L, and the difficulties of
length is far more exact than the regular ML exploration methods. The
experimental and numerical analyses, which take time, require exper-
Conditional Tabular Generative Adversarial Network (CTGAN) model,
tise, and cost a lot to find optimal parameters. They use ML for precise
which is the so-called "CT-SurpriseAF-BO," is instrumental in generating
density prediction to overcome these issues. The study uses ANN, KNN,
synthetic data. This novelty promotes the accuracy of prognostications
SVM, and linear regression techniques. Validation with R-squared values
by not having to engage in additional physical experiments. The study
and error metrics reveals improved performance for ANN and SVM (R² =
deliberately focusing on AI and ML technologies in cyber
0.95 and 0.923, respectively). ML models optimise the process param-
manufacturing, and especially, in MAM [91].
eters and might improve AM applications [51].
Rojek et al. (2024) illustrate the utilization of AI, such as ML, to
Christou et al. (2022) stress out that according to the report of Cancer
manage and develop AM processes. Certain duties are executed expertly
Conclusion, Hepatocellular carcinoma (HCC), that hails from the skies as
by this approach. AI enables expert operators to construct and supervise

Fig. 18. The radar plot with metrics findings for printability [89].

16
S. Inayathullah and R. Buddala Results in Engineering 25 (2025) 103676

autonomously acquired procedures in AM. The reduction of human printing conditions [10]. Additionally, these algorithms help with
error is achieved through the automation of product and material pressure detection and temperature characterization, assuring constant
manufacturing. AI in 3D technology enables ML computers with vision quality and performance. ML can help manufacturers make sense of
systems to analyze data like humans and create 3D objects. An incre- real-time data from sensors, understanding how temperature and speed
mental printer prints many materials at once, resulting in improved or the composition of materials lead to quality final products. This en-
quality. The method is rapid and precise. The model’s 97.56 % accuracy hances reliability and stability that enables them to provide a better
indicates precision and low inaccuracy. The AI-powered 3D printing turnaround time [96]. ML allows manufacturers to know the property of
device can analyse the surface and adjust the printing process based on materials they are handling and produce better AM parts that make the
material properties. Extending learning sets enables AI/ML-based sys- products more reliable at functional scales. A summary of the literature
tems like ours to adapt to various techniques, substances, or multi- reviewed in ML Methods for Characterization is shown in Table 8.
material printing. They allow production-line-specific ML solutions, Li et al. (2019) provides a data-driven predictive modelling strategy
such as self-learning systems, to adapt to production [92]. intended to improve surface integrity in AM. A variety of sensors are
Yan et al. (2024) emphasise the need of determining the purity of employed to gather data on vibration and temperature, while an
metal powder in order to guarantee the quality of additive-produced ensemble learning algorithm is utilized to forecast surface abnormal-
items. Contamination can cause fractures and problems with ities. Using time and frequency domains to extract and select charac-
completed products. Traditional testing methods are time-consuming teristics improves computational effectiveness and precision. The
and emphasise physical integrity above material composition. Hyper- algorithm accurately predicts the surface texture of 3D printed compo-
spectral imaging (HSI), when paired with ML, provides a nondestructive nents, and it has been confirmed using FFF testing [106].
solution by detecting tiny variations in temperature, moisture, and In their groundbreaking study, Zhu et al. (2021) provide the first use
chemical composition. This study investigates the use of HSI cameras, of PINN for metal AM melt pool temperature and fluid dynamics fore-
both near-infrared and visible, for nondestructive examination of metal casting. He utilized the PINN approach to address two typical metal
powders. It discusses technological issues and answers through three production issues. Fig. 19 depicts the fully linked deep neural network
case studies: creating a spectrum dictionary, identifying contamination, for metal AM. PINN’s early success in modeling and predicting complex
and analyzing band selection. The experimental findings show that HSI metal AM processes sets the door for its widespread implementation in
and ML may be used for nondestructive testing in powder metallurgy advanced manufacturing [49].
while addressing industrial production requirements [48]. Metamaterials, also called architected or programmable materials-
Hamdy (2024) uses the infinite objective AI for aesthetic dentistry, a according to Garland et al. (2021) allow designers to develop meta-
new way in his article. It is the digital diagnosis, prognosis, and the materials with freely shaped structures that would be almost impossible
treatment prediction that the AI is really good at. Despite being a young to occur naturally, and manipulate of mesoscale geometries to achieve
field, AI has taken the dental specialty fields such as radiography, end- material properties not found in nature. The gradual usage of AM in
odontics, orthodontics, implantology, prosthetics, and caries detection, different areas has made it possible to create complicated geometric
to a whole new level, of which almost every doctor in these fields relies metamaterials. However, creating new lattice metamaterials with
on it. The utilization of AI in dentistry may be categorized into four main complex features such as multi-physics interactions for high-
areas: evaluations, decision-making processes, scheduling therapy, and performance applications remains a tough task. Usually, designs are
outcome prognosis. Advanced dental applications encompass several based on human intuition or gradient optimization for simple problems.
cutting-edge technologies such as precise digital smile designs, AI- Garland et al. (2021) suggest that ML can discover new unit cells that
driven 3D printing, assessment of complex endodontic challenges, optimize both the elastic stiffness under a static load and the wave
finite element analysis (FEA), computer-aided design and computer- speed, which becomes a key factor during impact tests. These tests were
aided manufacturing (CAD/CAM), dental phonetic applications, and accomplished with relatively fewer simulation calls (3500) [47].
eye-tracking. To carry out the improvement in the quality of care that Silva et al. (2022) state that three-dimensional printing can produce
will up the efficiency, the doctor must be regularly updated about the simple and complex components from computerized designs using cost-
new trends and technologies in the field of AI [95]. effective materials and minimal waste. The article highlights AI’s
growing importance in ML and other fields. Silva and colleagues con-
3.9. Characterization in AM structed a neuroevolutionary model that predicts the tensile strength of
FFF 3D-printed materials using neural networks and genetic algorithms.
ML techniques are crucial in the characterization of AM processes. The Python model was trained with specific parameters and analyzed
They allow for extensive material characterization by analyzing data to
forecast and optimize attributes including tensile strength, impact Table 8
strength, compressive and flexural strength. Additionally, these algo- Summary of ML Methods for Characterization.
rithms help with pressure detection and temperature characterization, S. AM application ML technique used Reference
assuring constant quality and performance. By use of ML, manufacturers No
can have a good grasp of the characteristics of materials and this leads to
1 Detect temperature and Pool PINN [49]
better AM outcomes and the production of more dependable high- dynamics in Metal AM.
quality products. Table 6 presents a summary of the literature covered 2 Production and characterization ResNet, ConvNeXt, ViT [97]
in ML Methods for Characterization. and SWIN
ML techniques play a crucial role in characterizing AM processes. 3 Characterization KNN, DT [98]
4 Impact strength measurement GA and CNN [47]
Analysing the complex interactive model of AM and its properties now 5 Pressure detection Edge AI module [99]
permits characterization to be supported within AM on a condition- 6 Tensile strength measurement GA [100]
based basis with increased accuracy and speed. They also enable 7 Pressure detection SVM, ANN & KNN [101]
comprehensive material characterization via data analysis and provide 8 Impact strength measurement ANN and NSAI [102]
9 Pressure detection MOGOA and ANN [103]
robust fitting tools to predict and optimize attributes like tensile
10 Compressive and flexural SVMR [50]
strength, impact strength, compressive strength, and flexural strength. strengths measurement
Moreover, ML algorithms can process large datasets generated by AM to 11 Characterization Generative Convergence [104]
identify patterns and correlations between input parameters that con- Model (GCM)
ventional methods might miss, resulting in predictive models with 12 Tensile and Flexural strength LightGBM, H2OAutoML [105]
measurement
higher accuracy in predicting mechanical performance under specific

17
S. Inayathullah and R. Buddala Results in Engineering 25 (2025) 103676

Fig. 19. A deep neural network that is entirely linked and designed for metal AM [49].

using data and visuals. The model’s determination coefficient exceeded integrates additive printing, ML, and graphene nanoplatelets-based
90 %, demonstrating great accuracy and outperforming the tests [100]. piezoresistive foam to provide effective and adaptable pressure detec-
Pazienza et al. (2022) introduces the very advanced 3D-printed tion for many appliance, with significant advantages. The manufactured
Monitoring Equipment Mask Environment with an adaptive filter. foam’s mechanical response and strong piezoresistive sensitivities (0.31
Fig. 20 depicts design pictures of the (ME)2 3D-printed novel gas mask. kPa-1 at 8 kPa) are experimentally evaluated. Pressure sensing is
Critical sensors, Bluetooth, a system-on-a-chip micro-controller, and a accomplished by the use of electrical resistance tomography (ERT) uti-
rechargeable battery to monitor oxygen saturation, heart rate and body lizing the opposing current injection method. The gathered data is then
temperature, in the mask. Safe and effective health data monitoring analyzed using ML cataloging algorithms to determine the specific
involves strategically placed sensors and a smartphone app. Early location of pressure applied on the surface of the foam mat sensor. The
detection of respiratory and cardiovascular issues using edge AI modules experimental findings show that the method detects pressure across
allows predictive analysis, warnings, and suggestions. In synthetic data huge surfaces with an accuracy of 87.5 % or 83.7 % for distortion
validation testing, an ML model that used vital signs to distinguish generated by a single fingertip touch or two fingers touching different
COVID-19 from seasonal flu showed 94.80 % accuracy [99]. points on different FMS regions [101].
Neelam et al. (2022) predicts the mechanical properties of 3D prin- Lizarazu et al. (2023) devised an ML framework to precisely forecast
ted closed cell foams made of HDPE with different contents of hollow stress-strain correlations in arc-direct energy deposited mild steel.
glass microspheres infused therein and tensile and flexural strength of Approximately 1000 parameter sets were constructed via the Latin
these foams using automated and ensemble ML algorithms. The re- Hypercube Sampling Method, with synthetic Representative Volume
searchers analyze the available data pertaining to the mechanical Elements produced by Voronoi Tessellation. The input data, consisting
properties of these foams to try and determine the most effective ML of parameter sets and RVE pictures, underwent preprocessing tech-
algorithms that are proper for this intricate three-phased microstructure. niques such as standardisation, feature selection, scaling, and refining of
Through training with input parameters of orifice temperature, bed stress-strain curves by cubic splines and PCA. Two prediction models
temperature, and force values, experimental data was used to validate were utilized: one grounded in microstructural characteristics and the
strength predictions from the models. In comparison, LightGBM ranked other in RVE pictures, attaining a high accuracy with an RMSE of around
the best compared to other ensemble-based algorithms in the models 5 MPa (about 1 % of yield stress). These findings validate the efficacy of
that were evaluated; in turn, the H2OAutoML proved to be better ML in the accurate and efficient characterization of dual-phase steels
overall. Finally, the study also evaluates the memory that each model [105].
uses as well as the time consumed, whereby all the ML algorithms Geng et al. (2023) examine the application of ML in the field of
achieved accuracy rates of above 95 % [107]. construction 3D printing, with a specific emphasis on quality inspection,
Data scarcity in domains such as chemistry, pharmacology, and process control, and material design. In addition, he evaluates the ca-
materials science impedes AI applications owing to costly and secretive pacity of ML to improve sustainability, intellect, and efficiency in the
data collecting techniques. Wilkinson et al. (2023) solve the difficulty by building construction sector. His analysis provides unique perspectives
customizing a 3D printer to do autonomous, low-cost sample charac- on the advancement of construction technologies and the optimization
terisation with light microscopy. The approach, when combined with of workflows. ML, the most promising AI simulation, use computer
DL, enhances throughput while eliminating subjective biases in labeling, commands to anticipate and predict real-world outcomes. Izadgoshasb
increasing the study of pharmaceutical crystal morphology and allowing et al. (2021) attempt to determine and estimate the compressive prop-
for more access to data-driven research [97]. Pesce et al. (2023) erty of extruded 3DP concrete or mortar made of cement using a variety

Fig. 20. Assembly photographs of the (ME)2 3D-printed revolutionary mask [99].

18
S. Inayathullah and R. Buddala Results in Engineering 25 (2025) 103676

of methods. MOGOA and ANN are employed to conduct the study. He

integrates both approaches via the ANNMOGOA strategy to forecast the
compression strength of 3D-printed concrete, as ML accuracy is con-
strained by the number of data recordings, which is just a few years of
research for concrete 3D printing. Iteration produces good results [103].
Mishra et al. (2023) observed that the Neurosymbolic model accu-
rately predicts the impact strength of polylactic acid additives. A Neu-
rosymbolic model was developed by employing a NN structure
consisting of input, a pair of concealed layers, and the final output layer.
Additionally, a DT regressor (DTR) was utilized for emblematic repre-
sentation. R-squared (R2) and Mean Squared Error (MSE) values from
the experimental and evaluation datasets were used to compare the
model to a Simple ANN model. Fig. 21 depicts the NSAI framework used
for the study. This novel Neurosymbolic technique for assessing the
impact strength of PLA components in AM has great promise for
enhancing the process [102].
Over the past decade, concrete 3D printing technology has evolved, Fig. 22. SVMR to predict 3D-printed concrete specimen compressive and
but eco-friendly versions with fine aggregate substitutes remain under- flexural strengths [50].
studied. Zhu et al. (2024) create 25 eco-friendly 3D-printed concrete
mixes with fine particles, SF, Ru, and glass fiber. Simulations verify They develop a real-time porosity prediction method that is based on the
compressive and flexural strengths. As shown in Fig. 22, the study melt pool boundary characteristics that are derived through functional
employed SVMR method to forecast the compressive and flexural principal component analysis (FPCA) and heat monitoring. Supervised
strengths of the concrete specimens created by 3D printing. Compressive learning algorithms categorise melt pools to predict porosity, with KNN
strength diminishes and printing defects increase with rubber content. having the best accuracy (98.44 %) and Dt having the fewest false
Reduced compressive strength with increased fiber volume. Rubber and positives (0.03 %). The study shows that the morphological model
2.5 % fiber volume strengthen hybridized concrete. SF, GF, and Ru in- paired with supervised learning outperforms basic measures and may be
crease cost efficiency and mechanical performance, making in-situ used in different AM procedures [98].
concrete an eco-friendly choice [50]. Zheng et al. (2024) created an ML framework for helium bubble
Mohseni et al. (2024) investigate the influence of melt pool variables image analysis that uses the GCM to improve unsupervised segmenta-
on the production of defects, specifically porosity, in AM components. tion and generalization for nuclear material characterization. The GCM

Fig. 21. Framework of NSAI adopted [102].

19
S. Inayathullah and R. Buddala Results in Engineering 25 (2025) 103676

surpasses previously employed methods for severing helium bubbles, AM process is becoming an increasingly critical challenge. AM, addi-
predominantly those with uncertain confines, by integrating a gradient tionally referred to as 3D printing, brings forth a lot of advantages in
convergence layer with a generative model. It delivers exact measure- terms of flexibility, customization, and material efficiency. However, the
ments of bubble size, density, and quantity, displaying robustness across complexity of any AM process, either due to infill defects or any other
many datasets and situations, guaranteeing the trustworthy and profi- anomaly, may develop imperfections that will jeopardize product
cient interpretation of helium bubble pictures in the characterisation of quality and performance [67–69]. To solve this issue, the researchers
nuclear materials. The suggested method for teaching GCM ML models is have studied ML-based anomaly detection techniques for automated
shown in Fig. 23 [104]. detection of defects. ML anomaly detection techniques show significant
The ML approaches have been influential in the material character- promise in detecting various defects such as surface irregularities,
isation and process optimization in AM, but as technology progresses the dimensional deviations, and internal flaws. One of the big challenges for
approach to data integration also needs to improve. Lastly, it would be this subject is getting enough annotated data to train ML models. It is a
future work to incorporate additional forms of multi-modal data sources tough task to acquire a dataset capable of accurately representing
(including thermal and acoustic signals) into ML models for better AM defective samples [70–73]. This is due to the low occurrence of defects
material characterisation predictions. This integrated approach can lead in production. Thus, it makes it difficult to develop defect detection
to more stable and accurate forecasts for everything from improved AM algorithms that are both reliable and accurate. In this context, re-
products to higher quality materials. searchers followed both semi-supervised and unsupervised learning
approaches to improve the fault identification accuracy of unlabeled
4. Challenges and opportunities data [53,74,75]. Besides, DL advancements have been associated with
some advanced methodologies lately, which involve normalizing flows
In response to the increasing market demand for higher quality and that recently show promise for overcoming the problem of scant labeled
efficiency in products and processes, companies are compelled to devise data [108].
novel and innovative methods to optimise their production [18]. ML is
transforming how we enhance productivity, dependability, and adapt- 4.3. Process control and monitoring of AM
ability in 3D printing, a rapidly changing technology. This research in-
vestigates the intriguing possibilities for combining ML with 3D Because of the complexity of AM processes and the susceptibility of
printing, with an emphasis on enhancing finishing touches, optimizing 3D-printed items to manufacturing factors, dependable control and
designs, and monitoring and regulating the printing process. monitoring systems are important. ML has emerged as a critical tech-
nique for solving this dilemma [60,88]. It enables the design of
4.1. Design optimization of AM closed-loop control systems that can dynamically modify process pa-
rameters in real time. This capacity guarantees that the quality of
In the field of AM, component design is a vital step where ML may printed components is consistent throughout manufacturing, increasing
have a substantial influence. The intricacy and versatility of 3D printing overall dependability and efficiency [85–89]. Researchers have
necessitate a thorough, data-driven design process. Researchers utilized demonstrated that by implementing ML-driven sensors and feedback
ML algorithms to optimize the design of 3D printed components, taking loops, they can detect and correct deviations from desired process pa-
into account material properties, printing orientation, and part form. rameters like as temperature, layer thickness, and material flow in real
Researchers trained their algorithms on large datasets of both successful time during printing [109,110]. This method not only improves the
and unsuccessful 3D print efforts. This method has produced predictive reliability and uniformity of 3D-printed components, but also reduces
models that can forecast how 3D-printed components will behave, the need for substantial post-processing and rework. As a result, it im-
allowing designers to make better informed decisions and develop their proves the overall efficiency and cost-effectiveness of AM processes [48,
designs more efficiently [58–62]. This innovation has also ushered in the 90–92].
era of "generative design" methodologies, in which ML-powered algo-
rithms produce new part designs that are suited to specific performance 4.4. Post-processing and finishing in AM
requirements [64–66]. These improvements are opening the door for
game-changing advances in part optimization and invention [77, Finalizing and refining 3D-printed products after production used to
80–82]. be a laborious, error-prone, and labor-intensive manual procedure.
However, a change towards data-driven automation has occurred
4.2. Defects detection in AM recently, all because of ML. An exact prediction of the printed part’s
surface quality, dimensions, and mechanical properties can now be
With Industry 4.0 and smart manufacturing, defect detection in an made with the help of predictive models developed using ML

Fig. 23. Proposed method for GCM ML model training illustration [104].

20
S. Inayathullah and R. Buddala Results in Engineering 25 (2025) 103676

techniques. According to Grierson et al. (2021), this invention gives increased production, superior quality control levels, and efficient use of
manufacturers the ability to efficiently personalize their finishing pro- resources at different scales.
cedures [111]. Further, ML has played an important part in automating
how we choose and optimize post-processing variables like cutting 4.7. Model validation techniques
speeds, tool trajectories, and surface finishes. These modern procedures
not only increase the regularity and quality of completed products, but The validation should be used alongside cross-validation techniques
they also reduce the need for human labor, making the manufacturing like K-fold cross-validation and LOOCV (Leave One Out Cross Valida-
process more efficient [109,112,113]. tion) in the use of prediction accuracy to avoid overfitting of the model
and provide stable model performance. They are used in splitting the
4.5. Characterization in AM data into training and validation subsets iteratively for testing the model
across different samples within a dataset. All the given data is divided
As AM processes become more complex, making quality and design into equal sized segments or "folds" during K-fold cross-validation. The
flexibility paramount, it will be important for such capabilities to use fold appears once as the validation set, and all the remaining folds
advanced analytic tools like ML. The determination and understanding appear as training. Average performance metric obtained over
of the physical properties, including strength and thermal performance, numerous iterations serves as a measure of model performance with
of 3D-printed parts could be quite challenging. ML thus automates the reduced variance associated with random sampling. LOOCV is very
capture of this data, and the processing thereof further leads to efficient in most cases for small data sets because each point is valid
improved process monitoring and optimization [47,49,97–99]. Both once, leading to more accurate results. LOOCV is computationally
unsupervised and RL algorithms can independently learn from the expensive and inefficient, when it comes to larger sets of data, which are
manufacturing process, identify repeating patterns, and change them in too huge to be loaded into the memory all at once [114,115]. Such
such a way as to minimize faults and improve the quality of the product cross-validation strategies allow for better generalization by ML models
microstructure. and may even help to enhance the model’s predictive ability on new data
ML is going to make the life assessment of 3D-printed parts auto- for AM processes.
matic, allowing a much wider understanding of their environmental
impact and design improvements [50,100–103]. In addition, 4.8. Performance benchmarks of ML algorithms versus conventional
machine-learning models coupled with finite element analysis also made methods
multilayer composite materials with excellent mechanical features,
showcasing the versatility of this approach. As AM processes become Table 9 compares some of the AM ML algorithms applied in detail
more complex, making quality and design flexibility paramount, it will and focuses on the performance metrics relative to classical approaches.
be important for such capabilities to use advanced analytic tools like ML. The results show that the performance, efficiency, and scalability of ML
The determination and understanding of the physical properties, methodologies continue to improve. SVM attained a high area under the
including strength and thermal performance, of 3D-printed parts could curve (AUC) of 0.94, indicating dependable classification efficacy while
be quite challenging. ML thus automates the capture of this data, and the automating jobs that were once dependent on labor-intensive manual
processing thereof further leads to improved process monitoring and inspection. GNNs attained over 90 % prediction accuracy, notably
optimization [40,42,88–90]. Both unsupervised and RL algorithms can excelling with smaller labeled datasets, so addressing the limits of
independently learn from the manufacturing process, identify repeating conventional methods reliant on solving boundary value problems with
patterns, and change them in such a way as to minimize faults and established border conditions. Advanced neural networks, like CNN,
improve the quality of the product microstructure. Machine-learning exhibit resilience in high-dimensional design jobs and fault identifica-
models coupled with finite element analysis have also made it possible tion in intricate printing processes. CNN achieved 98.9 % accuracy in
to create multilayer composite materials with excellent mechanical identifying layer flaws in metal powder bed fusion (PBF), diminishing
properties, showcasing the versatility of this approach. Furthermore, dependence on manual inspection, which is susceptible to human
automated life assessment of 3D-printed parts will allow deeper insights mistake and expensive. In optimization challenges, ANNs combined
into their environmental impact and future design improvements [43, with GAs attained R² values over 90 % in predicting tensile strength,
91–94]. therefore simplifying the laborious manual tuning processes typical of
conventional trial-and-error approaches. Likewise, models such as MLP
4.6. Scalability of ML in AM surpassed conventional models in both training and testing stages,
achieving prediction accuracies greater than 95 %.
Scalability is an important factor with regard to the adoption of ML The ensemble approach, such as combining RF and XGBoost, also
in AM since this determines the feasibility of applying models of ML generates excellent predictive accuracy at 99.75 %, ensuring better
across different scales of production and operational settings. ML ap- reliability and interpretability in the diagnostics of AM problems.
plications currently developed for AM are mostly intended for specific Hybrid approaches allowed for fast processing of physical parameters;
datasets, machines, or materials and lack significant transferability. for example, the combination of CAEs with clustering algorithms
However, solutions scalable enough are given by approaches like TL, improved laser point distances, powder layer thickness, and scanning
wherein models trained on one machine or material can be transferred speed-operations that usually require the assessment of an expert and
or adapted to others, and thus significantly reduce the need for extensive manual changes. In addition, KNN and other neural network models
retraining. Cloud-based computing and distributed ML frameworks were used to obtain 87.5 % accuracy in pressure zones detection of smart
further make the technology applicable for real-time data processing materials 3D printed overcoming the limitations of discrete sensor-based
and updates of models across the different machines while allowing approaches with synchronization and issues of spatial resolution. The
centralized control and monitoring across distributed manufacturing chart indicates that the performance of ML models is superior to con-
setups. Besides, hybrid ML models, combining data-driven approaches ventional approaches in terms of prediction accuracy and offers
with physics-based frameworks, allow for scalable performance across tremendous reduction in time, man-power, and computer usage.
AM processes without requiring an entire overhaul of data for each new Traditional methods require much testing, computationally expensive
application. All of these strategies reduce the computational costs and simulation (such as finite element modeling), or require technical
training times of models while making ML solutions more responsive to expertise, whereas ML methodologies reduce time by automating se-
evolving production requirements. Tackling scalability issues will be the lection of parameters, betterment of material characteristics, and real-
way to unlock the vast industrial benefits of ML in AM, which include time fault identification. The advantages, along with the ability to

21
S. Inayathullah and R. Buddala Results in Engineering 25 (2025) 103676

Table 9 Table 9 (continued )

ML Algorithms Performance Benchmarks vs Conventional Techniques. Algorithms Used Measures of Algorithm Analogies with Reference
Algorithms Used Measures of Algorithm Analogies with Reference Performance Traditional
Performance Traditional Approaches
Approaches
RF, DL RF algorithm got the Conventional [81]
Hierarchical High accuracy due to Conventional [58] best accuracy, methods are empirical
Clustering, SVM SVM classifier trained methods rely on precision, and F1- in nature, using expert
with industrial designer’s personal score of 88.1 %, 90.6 knowledge along with
examples; flexible experience and %, and 87.0 %; DL had plenty of trial-and-
clustering without knowledge, which can a maximum of recall at error work.
preset clusters. be time-consuming 87.3 %.
and inefficient. Convolutional High-quality results Traditional parameter [78]
3D CNN Higher accuracy and Linear regression [59] Autoencoder reflected traditional optimization is
generalizability across models were less (CAE), KMeans parameter manual, time-
different design accurate and often Mini Batch optimization consuming, and relies
repositories; robust for overfitted, especially observations; heavily on human
high-dimensional with high- clustering and scoring judgment.
voxel-based regression dimensional data. matched trends for
tasks. optimal laser point
GNNs High prediction Involves addressing [60] distance, powder layer
accuracy (~90 %) boundary value issues thickness, and laser
after several (BVPs) using specified scanning speed.
generations; efficient boundary conditions Support Vector High accuracy in Conventional [48]
with small labeled (BCs), hence Machine (SVM), detecting methods focus on
datasets. restricting flexibility. RF, Neural contamination in physical integrity
CNNs, Point Attained an average Frequently [62] Network, metal powders; PRBS rather than material
Depth Models RMS error of 1.00 mm necessitates further Proximity effectively reduced composition and can
(PDMs), for femoral CT or MRI scans to Ranking-Based data redundancy be time-consuming.
Statistical reconstructions and enhance implant Band Selection while maintaining
Shape Models 1.07 mm for tibial compatibility, hence (PRBS) representative
(SSMs) reconstructions. elevating expenses information.
and radiation ANN, KNN, ANN and SVM models Traditional methods [51]
exposure. Support Vector performed very well include trial-and-
ML, Maxwell’s Attained a maximum Traditional [63] Machine (SVM), with an R-squared error, FEM, and DOE,
Reciprocal enhancement of 25 % approaches may not Linear value of 0.95 and which consume much
Diagrams in out-of-plane provide such levels of Regression (LR) 0.923 in density time and are
stiffness for 3D- optimization and prediction, computationally
printed wing efficiency respectively. expensive and require
structures. knowledge in
DL, TL Achieved accuracies of Traditional methods [66] materials science and
80 %, 95 %, and 96 % rely on manual mechanical
for training, application of engineering.
validation, and testing geometrical ANN, SVM, RF Optimized ML models Formulation [118]
in BO classification; constraints that lower predicted the development can be
decreased stressors by the precision and printability and an extremely time
50 %. increase work effort. mechanical features of consuming and
Linear MLP performed far The traditional [116] filaments with a 84 % iterative process that
Regression, better in classification method could not be accuracy of necessitates a
Support Vector tasks than other as effective and predictions. significant number of
Machine (SVM), methods; with both precise. Processing adjustments in the
RF, Multilayer the training and temperatures from composition of
Perceptron testing dataset, HME and FDM had filaments as well as
(MLP) average prediction mean absolute errors the printing
accuracy is >95 %. as high as 5.5 ◦ C and parameters.
Support Vector High AUC (0.94) for Manual inspection is [73] 8.4 ◦ C, respectively.
Machine (SVM), SVM; effective lagging in reliability Deep CNN Attained 98.9 % Conventional quality [86]
Density-Based clustering with and speed. accuracy in the control approaches in
Spatial DBSCAN. classification of 9280 PBFM are costly and
Clustering with previously time-consuming,
Application of unexamined layer necessitating human
Noise photographs for inspection and
(DBSCAN) quality assurance in susceptible to user-
RF, XGBoost, Prediction Accuracy For fault diagnosis, [72] metal powder bed induced
SHapley for ensemble learning the traditional fusion. misinterpretations.
Additive model XGBoost and methods might not ANN, SVM, RF, ANN and SVM models Conventional [105]
exPlanations RF, is 99.75 % after appear to be as CNN have predicted stress- techniques for stress-
(SHAP) the application of trustworthy or open- strain curves of strain curve
these models. book as they really additively determination are
are. manufactured mild extremely costly and
Self-Organizing For geometric Big geometric flaws [117] steel with an RMSE of time-consuming along
Map (SOM) correctness of the may be missed by around 5 MPa, that is with protracted
component, an traditional GD&T ~1 % of the yield experimental
unsupervised ML features; it will stress. The image- procedures CNN.
technique based on consume much human based CNN predictions
SOM required only 3 work. have demonstrated
% of >1 million data good convergence but
points. with bigger errors.
(continued on next page)

22
S. Inayathullah and R. Buddala Results in Engineering 25 (2025) 103676

Table 9 (continued ) impact on manufacturing.

Algorithms Used Measures of Algorithm Analogies with Reference
Performance Traditional 5. Conclusion
Approaches

ANN, Multi- The ANNMOGOA-1 Traditional methods [103] ML’s incorporation into several phases of AM has opened hitherto
Objective was the most accurate for estimation of unheard-of possibilities for improving dependability, efficiency, and
Grasshopper model of all, with an compressive strength adaptability in this fast-changing industry. Key areas where ML has had
Optimization RMSE of 4.49 MPa and of 3D-printed mortar
a major influence are design optimization, defect detection, post-
Algorithm MAPE of 8.75 %, while are less effective and
(MOGOA) the accuracy of require more processing and finishing, and process management and monitoring
ANNMOGOA-2 was experimental data. highlighted in this paper. These areas directly concern the goal of the
the worst, with an study in finding how ML aids in AM processes and solves its unique
RMSE of 18.69 MPa difficulties indeed.
and a MAPE of 52.40
%.
ML algorithms predict the properties and parameters of AM by using
NSAI, Simple The NSAI model The traditional [102] large data sets of historical process parameters, such as temperature,
ANN demonstrated a methods used for speed, and feed rates of different materials, plus their effects on the
reduced MSE (2.7448 estimating the impact output. These algorithms drive data models to predict the mechanical
for training, 2.7026 strength of PLA
properties, such as tensile strength and surface roughness, for inputs.
for validation) and components prepared
elevated R-squared through AM are less This allows for real time ML to set parameters dynamically during
values (0.9840 for accurate and fail to manufacture in response to current sensor data improving the control
training, 0.9850 for utilize the strengths of quality of the process and reduction of defects.
validation) in contrast both neural networks Predictive analysis up-scales the efficiency, reliability, and quality of
to the Simple ANN and symbolic AI
model (MSE: 3.4174
the process with respect to AM. ML techniques have made predictive
for training, 3.3666 models for surface roughness, dimensional correctness, and mechanical
for validation; R- qualities possible in post-processing and finishing. These models let
squared: 0.9800 for producers more precisely modify their post-processing techniques,
training, 0.9813 for
hence lowering variability and guaranteeing better end component
validation).
SVM, ANN, KNN The KNN algorithm Traditional pressure [101] quality. Automation of parameter selection and optimization has
attained the maximum detection techniques simplified procedures, reducing reliance on human expertise, improving
accuracy in employing discrete consistency, and fulfilling the goal of developing autonomous and effi-
identifying pressure multi-sensor cient manufacturing systems.
zones on a 3D-printed arrangements
graphene-based encounter constraints
"Generative design" consists of providing unique part designs based
piezoresistive foam in installation, on prediction models bound to specific performance requirements. This
mat, with roughly synchronization, data technology enables designers to consider a more extensive ‘space’ of
87.5 % accuracy for acquisition, possibilities to iterate efficiently, and to promote the generation of
single touch and 83.7 processing, spatial
future forward advancements in AM as articulated by the study. Addi-
% for double touch resolution, and
cases. coverage. tionally, validating results against real world applications is critical to
ANN, GA The Conventional [100] increase the accuracy of simulations in AM. That includes tweaking
neuroevolutionary methods for models to work with experimental data, tuning parameters to resolve
model, which predicting tensile differences and to compare predictions of properties’ classes against
combines artificial strength are
neural networks and ineffective and need
industry norms. A combination of cross validation across materials and
GAs, achieved a extensive processes, and iterative feedback, is demonstrated in increasing pre-
determination experimental data and diction accuracy and generalizing results across AM applications. These
coefficient (R²) of 90 manual adjustment. results directly answer the two objectives of the study: alleviating AM
%, indicating a high
deployment bottlenecks, and enhancing the utility of ML in AM.
degree of accuracy in
predicting the tensile Enrichment of the characterization of AM processes is achieved by
strength of 3D-printed hybrid data collection methods that combine real time sensor data with
components. historical datasets. By improving the predictive modeling accuracy,
these methods enable more advanced monitoring and control systems
that minimize post processing and boost quantitative production effi-
process large data sets at high speeds, indicate that ML frameworks are
ciency. This integration completes the vision of connecting
gradually replacing traditional techniques in many fundamental AM
manufacturing improvement with data driven insights. Real-time sensor
applications.
feedback and closed-loop control algorithms detect and correct de-
viations of specified process parameters; therefore, 3D-printed products
4.9. Industrial applications of ML in AM and tangible improvements exhibit improved dependability and repeatability. This work highlights
the fact that these advancements, enabled by ML, represent a game
Table 10 illustrates the industrial-scale ML applications for AM, changing paradigm shift in AM that makes it more scalable and cost
across different sectors. ML is being used for design optimization, defect effective compared to many other manufacturing technologies. At this
detection and production efficiency by companies such as General point, the alignment of ML and AM signals an exciting, innovative, and
Electric, BMW or Siemens. To do that, GE has reduced material waste by efficient future to look forward. In the hopes of revolutionizing the
up to 20 % in 3D-printed engine parts and BMW has cut production manufacturing landscape with ML, this collaboration will continue to
defects by as much as 30 % through real-time AI monitoring. In health catalyze the evolution of the industrial field, which will allow engineers
care, Stryker has achieved 20 % improvements in implant accuracy that and designers to reinvent what production capabilities are possible.
drives better patient results. They highlight the significant real-world
improvements that can be derived from coupling ML with AM to such Future recommendations
as waste reduction, higher quality, shorter time-to-print and improved
part performance, showing just how ML is making a demonstrable Future research must focus on these seven categories for greater

23
S. Inayathullah and R. Buddala Results in Engineering 25 (2025) 103676

Table 10
Industrial Applications of ML in AM and Tangible Improvements [119–124].
Sector Organization Application The enhancements Impact

Aerospace General ML is utilized to improve the design and 20 % reduction in material waste, 15 % increase Reduces material waste, improves
Electric (GE) predict the performance of 3D-printed engine in production speed, improved component component performance, and accelerates
components. strength. production timelines.
Automotive BMW AI oversees the quality of printed components 25 % decrease in manufacturing faults, 10 % Reduces faults, improves manufacturing
in real-time throughout the manufacturing increase in production speeds, reduced rework efficiency, and lowers costs related to
process. expenses. faulty components.
Medical Stryker AI-driven AM for the creation of customized Enhanced implant fit precision, 30 % reduction Improves the accuracy of implants,
implants for patients. in surgical revisions, 10 % expedited design-to- resulting in better patient outcomes and
manufacturing process. fewer surgical revisions.
Energy Siemens ML techniques are utilized for instantaneous 30 % drop in failure rates, enhanced Enhances component quality by detecting
flaw identification using sensor data. dependability of components, reduced and rectifying flaws during the 3D printing
downtime. procedure.
Consumer Adidas Optimization of design for 3D-printed Enhanced product customisation, 50 % Attains mass customisation while
Goods footwear, particularly midsoles, by AI. expedited prototyping, augmented material preserving durability and performance in
durability. sporting footwear.
Defense Lockheed AI is employed to enhance material selection 10 % decrease in component mass, improved Decreases component weight, augments
Martin and forecast structural performance in AM for material durability, accelerated design durability, and improves manufacturing
defense components. iterations. process efficiency.

deployment of ML in AM. By addressing these areas, the deployment of CRediT authorship contribution statement
ML in AM can be accelerated, overcoming current bottlenecks and
fostering future innovations. Sirajudeen Inayathullah: Writing – original draft, Methodology,
Investigation, Formal analysis, Data curation, Conceptualization. Rav-
1. Improving Data Availability and Standardization iteja Buddala: Writing – review & editing, Validation, Supervision.
Focus on establishing cross sector guidelines for standard data
collection and sharing. Applying best practices around data format- Declaration of competing interest
ting and quality will help enable the precision of services like AM
where machine learning algorithms written by service providers are We have no conflicts of interest to disclose.
not built with human variability in the interpretation of measure-
ments as a baked component. Data availability
2. Enhancing ML Algorithms
More sophisticated ML algorithms would be required for an ac- No data was used for the research described in the article.
curate prediction of material properties, allowing better simulation,
process optimization and reliability. All these algorithms must take References
into account not only material behavior, but the process parameters
as well to make holistic predictions. [1] I.G. Ian Gibson, Additive manufacturing technologies 3D printing, rapid
prototyping, and direct digital manufacturing, (2015). https://doi.org/10.100
3. AI-Driven Tool and Design Development 7/978-1-4939-2113-3.
Creating advanced AI supported designs and 3D printing tools [2] T.D. Ngo, A. Kashani, G. Imbalzano, K.T.Q. Nguyen, D. Hui, Additive
optimized to take full advantage of the diverse attributes custom- manufacturing (3D printing): a review of materials, methods, applications and
challenges, Compos. B Eng. 143 (2018) 172–196, https://doi.org/10.1016/j.
izable with AM (geometric flexibility, weight reduction). Utilizing AI compositesb.2018.02.012.
in Optimizing Custom Manufacturing Processes using technology to [3] F. Mashhadi, S.A.S. Monroy, Deep learning for optimal resource allocation in IoT-
optimize design parameters enabled additive manufacturing, in: 2020 IEEE 6th World Forum on Internet of
Things (WF-IoT), IEEE (2020) 1–6, https://doi.org/10.1109/WF-
4. Advanced Sensor Integration and Real-Time Monitoring IoT48130.2020.9221038.
Expenditure must be increased to support the creation of new [4] T. Tracy, L. Wu, X. Liu, S. Cheng, X. Li, 3D printing: innovative solutions for
generation sensor systems that would enable measurement of tem- patients and pharmaceutical industry, Int. J. Pharm. 631 (2023), https://doi.org/
10.1016/j.ijpharm.2022.122480.
perature, stress and other important features during printing.
[5] F. Han, S. Liu, J. Zou, Y. Ai, C. Xu, Defect detection: defect classification and
Enhanced real-time diagnostic technology will aid in defect deter- localization for additive manufacturing using deep learning method, in: 2020
mination, inspection of quality and process control. 21st International Conference on Electronic Packaging Technology (ICEPT), IEEE,
5. Overcoming the ‘Interpretability Barrier’ 2020, pp. 1–4, https://doi.org/10.1109/ICEPT50128.2020.9202566.
[6] R. Nascimento, I. Martins, T.A. Dutra, L. Moreira, Computer vision based quality
Interpretable and explainable ML models for AM, that would instill control for additive manufacturing parts, Int. J. Adv. Manuf. Technol. 124 (2023)
the confidence in decisions made using AI. Transparency in the ML 3241–3256, https://doi.org/10.1007/s00170-022-10683-5.
model’s predictions will help to report them across industries, with [7] R. Li, M. Jin, Z. Pei, D. Wang, Geometrical defect detection on additive
manufacturing parts with curvature feature and machine learning, Int. J. Adv.
more confidence, particularly when applied on the applications that Manuf. Technol. 120 (2022) 3719–3729, https://doi.org/10.1007/s00170-022-
have a more impact. 08973-z.
6. Interdisciplinary Collaboration [8] A. García-Collado, J.M. Blanco, M.K. Gupta, R. Dorado-Vicente, Advances in
polymers based Multi-Material Additive-Manufacturing Techniques: state-of-art
Speed up innovations and tackle major issues by promoting more review on properties and applications, Addit. Manuf. 50 (2022) 102577, https://
interaction between governments, industries as well as universities. doi.org/10.1016/j.addma.2021.102577.
These partnerships will help to faster implement ML and comple- [9] J. Liu, J. Ye, D. Silva Izquierdo, A. Vinel, N. Shamsaei, S. Shao, A review of
machine learning techniques for process and performance optimization in laser
ment the sharing of best practice as well as data. beam powder bed fusion additive manufacturing, J. Intell. Manuf. 34 (2023)
7. Developing Workforce Competence 3249–3275, https://doi.org/10.1007/s10845-022-02012-0.
Additional training and education programs should be completely [10] Z. Wang, W. Yang, Q. Liu, Y. Zhao, P. Liu, D. Wu, M. Banu, L. Chen, Data-driven
modeling of process, structure and property in additive manufacturing: a review
invested in ensuring that a well staff force is in place to use ML in AM.
and future directions, J. Manuf. Process. 77 (2022) 13–31, https://doi.org/
Widespread and successful deployment of the technologies is crucial 10.1016/j.jmapro.2022.02.053.
in this context.

24
S. Inayathullah and R. Buddala Results in Engineering 25 (2025) 103676

[11] R. Ranjan, R. Samant, S. Anand, Integration of design for manufacturing methods heat exchanger using machine learning model and fire hawk optimizer, Case
with topology optimization in additive manufacturing, J. Manuf. Sci. Eng. 139 Stud. Therm. Eng. 49 (2023) 103294, https://doi.org/10.1016/j.
(2017) 061007, https://doi.org/10.1115/1.4035216. csite.2023.103294.
[12] B.R. Hunde, A.D. Woldeyohannes, Future prospects of computer-aided design [34] F. Djouider, M. Abd Elaziz, A. Alhawsawi, E. Banoqitah, E.B. Moustafa, A.
(CAD)–A review from the perspective of artificial intelligence (AI), extended H. Elsheikh, Experimental investigation and machine learning modeling using
reality, and 3D printing, Results. Eng. 14 (2022) 100478, https://doi.org/ LSTM and special relativity search of friction stir processed AA2024/Al2O3
10.1016/j.rineng.2022.100478. nanocomposites, J. Mater. Res. Technol. 27 (2023) 7442–7456, https://doi.org/
[13] M. Attaran, The rise of 3-D printing: the advantages of additive manufacturing 10.1016/j.jmrt.2023.11.155.
over traditional manufacturing, Bus. Horiz. 60 (2017) 677–688, https://doi.org/ [35] W. Rajhi, A.M.M. Ibrahim, A.-H.I. Mourad, M. Boujelbene, M. Fujii, A. Elsheikh,
10.1016/j.bushor.2017.05.011. Prediction of milled surface characteristics of carbon fiber-reinforced
[14] J.H. KS, M.P. Subramaniam, M.K. KT, P. Sreeram, S. Parvathi, P.S. Sari, polyetheretherketone using an optimized machine learning model by gazelle
A. Pullanchiyodan, D.M. Mulhivill, P. Raghavan, Fabrication and challenges of 3D optimizer, Measurement 222 (2023) 113627, https://doi.org/10.1016/j.
printed sensors for biomedical applications-Comprehensive review, Results. Eng. measurement.2023.113627.
(2024) 101867, https://doi.org/10.1016/j.rineng.2024.101867. [36] A. Muqeet, A. Israr, M.H. Zafar, M. Mansoor, N. Akhtar, A novel optimization
[15] T.C. Dzogbewu, N. Amoah, S.A. Jnr, S.K. Fianko, D.J. de Beer, Multi-material algorithm based PID controller design for real-time optimization of cutting depth
additive manufacturing of electronics components: a bibliometric analysis, and surface roughness in finish hard turning processes, Results. Eng. 18 (2023)
Results. Eng. 19 (2023) 101318, https://doi.org/10.1016/j.rineng.2023.101318. 101142, https://doi.org/10.1016/j.rineng.2023.101142.
[16] Y. Li, C. Su, J. Zhu, Comprehensive review of wire arc additive manufacturing: [37] H.K. Elminir, M.A. El-Brawany, D.A. Ibrahim, H.M. Elattar, E.A. Ramadan, An
hardware system, physical process, monitoring, property characterization, efficient deep learning prognostic model for remaining useful life estimation of
application and future prospects, Results. Eng. 13 (2022) 100330, https://doi. high speed CNC milling machine cutters, Results. Eng. 103420 (2024), https://
org/10.1016/j.rineng.2021.100330. doi.org/10.1016/j.rineng.2024.103420.
[17] M. Ali, U. Sajjad, I. Hussain, N. Abbas, H.M. Ali, W.-M. Yan, C.-C. Wang, On the [38] A. Harmayanti, I.P. Tama, F. Gapsari, Z. Akbar, H. Juliano, Cardiac biometrics
assessment of the mechanical properties of additively manufactured lattice and perceived workload regression analysis using random forest regressor in
structures, Eng. Anal. Bound. Elem. 142 (2022) 93–116, https://doi.org/ cognitive manufacturing tasks, Int. J. Mech. Eng. Technol. Applic. 5 (2024)
10.1016/j.enganabound.2022.05.019. 108–120, https://doi.org/10.21776/MECHTA.2024.005.01.11.
[18] J. Mazumder, Design for metallic additive manufacturing machine with [39] C. Wang, X.P. Tan, S.B. Tor, C.S. Lim, Machine learning in additive
capability for “Certify as You Build, Procedia CIRP. 36 (2015) 187–192, https:// manufacturing: state-of-the-art and perspectives, Addit. Manuf. 36 (2020)
doi.org/10.1016/j.procir.2015.01.009. 101538, https://doi.org/10.1016/j.addma.2020.101538.
[19] A. Boretti, A perspective on 3D printing in the medical field, Annals of 3D Printed [40] M. Danish, M.K. Gupta, S.A. Irfan, S.M. Ghazali, M.F. Rathore, G.M. Krolczyk,
Medicine 13 (2024) 100138, https://doi.org/10.1016/j.stlm.2023.100138. A. Alsaady, Machine learning models for prediction and classification of tool wear
[20] H. Li, Machine-Learning Based Modelling for AM Processes, in: Proceedings of the in sustainable milling of additively manufactured 316 stainless steel, Results. Eng.
2020 IEEE 10th International Conference on “Nanomaterials: Applications and 22 (2024), https://doi.org/10.1016/j.rineng.2024.102015.
Properties” 2020, NAP, 2020, pp. 1–4, https://doi.org/10.1109/ [41] A.A. Esoso, O.M. Ikumapayi, T.-C. Jen, E.T. Akinlabi, Exploring Machine Learning
NAP51477.2020.9309628. Tools for Enhancing Additive Manufacturing: a Comparative Study, Ingenierie
[21] X. Li, X. Jia, Q. Yang, J. Lee, Quality analysis in metal additive manufacturing Des Systemes d’Information 28 (2023) 535–544, https://doi.org/10.18280/
with deep learning, J. Intell. Manuf. 31 (2020) 2003–2017, https://doi.org/ isi.280301.
10.1007/s10845-020-01549-2. [42] C. Hyunjin, A study on the change of manufacturing design process due to the
[22] T. Wuest, D. Weimer, C. Irgens, K.D. Thoben, Machine learning in manufacturing: development of AI design and 3D printing, in: IOP Conf Ser Mater Sci Eng, IOP
advantages, challenges, and applications, Prod. Manuf. Res. 4 (2016) 23–45, Publishing, 2020 012010, https://doi.org/10.1088/1757-899X/727/1/012010.
https://doi.org/10.1080/21693277.2016.1192517. [43] Y. Xiong, Y. Tang, Q. Zhou, Y. Ma, D.W. Rosen, Intelligent additive manufacturing
[23] F. Mashhadi, S.A. Salinas Monroy, Deep learning for optimal resource allocation and design: state of the art and future perspectives, Addit. Manuf. 59 (2022)
in IoT-enabled additive manufacturing, in: IEEE World Forum on Internet of 103139, https://doi.org/10.1016/j.addma.2022.103139.
Things, WF-IoT 2020 - Symposium Proceedings, 2020, https://doi.org/10.1109/ [44] J. Wang, Y. Ma, L. Zhang, R.X. Gao, D. Wu, Deep learning for smart
WF-IoT48130.2020.9221038. manufacturing: methods and applications, J. Manuf. Syst. 48 (2018) 144–156,
[24] C.M. Thakar, R.J.M. Ventayen, E.H. Ramirez-Asis, J.E. Vilchez-Carcamo, M. https://doi.org/10.1016/j.jmsy.2018.01.003.
E. Maguiña-Palma, A. Thommandru, An Investigation on Impact of Machine [45] S.L. Chan, Y. Lu, Y. Wang, Data-driven cost estimation for additive manufacturing
Learning in Additive Manufacturing, in: Proceedings of Second International in cybermanufacturing, J. Manuf. Syst. 46 (2018) 115–126, https://doi.org/
Conference in Mechanical and Energy Technology: ICMET 2021, India, Springer, 10.1016/j.jmsy.2017.12.001.
2022, pp. 307–313, https://doi.org/10.1007/978-981-19-0108-9_32. [46] M.S. Meiabadi, M. Moradi, M. Karamimoghadam, S. Ardabili, M. Bodaghi,
[25] E. Hirata, D. Watanabe, A. Chalmoukis, M. Lambrou, A topic modeling approach M. Shokri, A.H. Mosavi, Modeling the producibility of 3D printing in polylactic
to determine supply chain management priorities enabled by digital twin acid using artificial neural networks and fused filament fabrication, Polymers.
technology, Sustainability (Switzerland) 16 (2024), https://doi.org/10.3390/ (Basel) 13 (2021) 3219, https://doi.org/10.3390/polym13193219.
su16093552. [47] A.P. Garland, B.C. White, S.C. Jensen, B.L. Boyce, Pragmatic generative
[26] A.M. Alhawsawi, E.B. Moustafa, M. Fujii, E.M. Banoqitah, A. Elsheikh, Kerf optimization of novel structural lattice metamaterials with machine learning,
characteristics during CO2 laser cutting of polymeric materials: experimental Mater. Des. 203 (2021), https://doi.org/10.1016/j.matdes.2021.109632.
investigation and machine learning-based prediction, Engineering Science and [48] Y. Yan, J. Ren, H. Sun, R. Williams, Nondestructive quantitative measurement for
Technology, an Int. J. 46 (2023) 101519, https://doi.org/10.1016/j. precision quality control in additive manufacturing using hyperspectral imagery
jestch.2023.101519. and machine learning, IEEe Trans. Industr. Inform. (2024), https://doi.org/
[27] E.M.S. El-Said, M. Abd Elaziz, A.H. Elsheikh, Machine learning algorithms for 10.1109/TII.2024.3384609.
improving the prediction of air injection effect on the thermohydraulic [49] Q. Zhu, Z. Liu, J. Yan, Machine learning for metal additive manufacturing:
performance of shell and tube heat exchanger, Appl. Therm. Eng. 185 (2021) predicting temperature and melt pool fluid dynamics using physics-informed
116471, https://doi.org/10.1016/j.applthermaleng.2020.116471. neural networks, Comput. Mech. 67 (2021) 619–635, https://doi.org/10.1007/
[28] A.H. Elsheikh, Applications of machine learning in friction stir welding: s00466-020-01952-9.
prediction of joint properties, real-time control and tool failure diagnosis, Eng. [50] R. Zhu, K.-J.I. Egbe, H. Salehi, Z. Shi, P. Jiao, Eco-friendly 3D printed concrete
Appl. Artif. Intell. 121 (2023) 105961, https://doi.org/10.1016/j. with fine aggregate replacements: fabrication, characterization and machine
engappai.2023.105961. learning prediction, Constr. Build. Mater. 413 (2024), https://doi.org/10.1016/j.
[29] S. Shanmugan, A.H. Elsheikh, R. Sathyamurthy, A.K. Thakur, M. Issa, H. Panchal, conbuildmat.2024.134905.
T. Muthuramalingam, R. Kumar, M. Sharifpur, Low-cost bilayered structure for [51] M. Gor, A. Dobriyal, V. Wankhede, P. Sahlot, K. Grzelak, J. Kluczyński,
improving the performance of solar stills: performance/cost analysis and water J. Łuszczek, Density Prediction in Powder Bed Fusion Additive Manufacturing:
yield prediction using machine learning, Sustain. Energy Technol. Assessments 49 machine Learning-Based Techniques, Applied Sciences (Switzerland) 12 (2022),
(2022) 101783, https://doi.org/10.1016/j.seta.2021.101783. https://doi.org/10.3390/app12147271.
[30] T.K. Devi, E.B. Priyanka, P. Sakthivel, Paper quality enhancement and model [52] A. Nyabadza, L.M. Mc Donough, A. Manikandan, A.B. Ray, A. Plouze,
prediction using machine learning techniques, Results. Eng. 17 (2023) 100950, C. Muilwijk, B. Freeland, M. Vazquez, D. Brabazon, Mechanical and antibacterial
https://doi.org/10.1016/j.rineng.2023.100950. properties of FDM additively manufactured PLA parts, Results. Eng. 21 (2024)
[31] A. Elsheikh, M. Zayed, A. Aboghazala, F.A. Essa, S. Rehman, O.L. Muskens, 101744, https://doi.org/10.1016/j.rineng.2023.101744.
A. Kamal, M.Abd Elaziz, Innovative solar distillation system with prismatic [53] M. Danish, M.K. Gupta, S.A. Irfan, S.M. Ghazali, M.F. Rathore, G.M. Krolczyk,
absorber basin: experimental analysis and LSTM machine learning modeling A. Alsaady, Machine learning models for prediction and classification of tool wear
coupled with great wall construction algorithm, Process Safe. Environ. Protect. in sustainable milling of additively manufactured 316 stainless steel, Results. Eng.
186 (2024) 1120–1133, https://doi.org/10.1016/j.psep.2024.04.063. 22 (2024) 102015, https://doi.org/10.1016/j.rineng.2024.102015.
[32] M. Abd Elaziz, M.E. Zayed, H. Abdelfattah, A.O. Aseeri, E.M. Tag-eldin, M. Fujii, [54] G.D. Goh, S.L. Sing, W.Y. Yeong, A review on machine learning in 3D printing:
A.H. Elsheikh, Machine learning-aided modeling for predicting freshwater applications, potential, and challenges, Artif. Intell. Rev. 54 (2021) 63–94,
production of a membrane desalination system: a long-short-term memory https://doi.org/10.1007/s10462-020-09876-9.
coupled with election-based optimizer, Alexandria Eng. J. 86 (2024) 690–703, [55] I.Z. Era, M.A. Farahani, T. Wuest, Z. Liu, Machine learning in Directed Energy
https://doi.org/10.1016/j.aej.2023.12.012. Deposition (DED) additive manufacturing: a state-of-the-art review, Manuf. Lett.
[33] M. Mudhsh, E.M.S. El-Said, A.O. Aseeri, R. Almodfer, M. Abd Elaziz, S. 35 (2023) 689–700, https://doi.org/10.1016/j.mfglet.2023.08.079.
M. Elshamy, A.H. Elsheikh, Modelling of thermo-hydraulic behavior of a helical

25
S. Inayathullah and R. Buddala Results in Engineering 25 (2025) 103676

[56] T. Wuest, C. Irgens, K.-D. Thoben, An approach to monitoring quality in [81] H. Chen, Y. Liu, S. Balabani, R. Hirayama, J. Huang, Machine Learning in
manufacturing using supervised machine learning on product state data, J. Intell. Predicting Printable Biomaterial Formulations for Direct Ink Writing, Research. 6
Manuf. 25 (2014) 1167–1180, https://doi.org/10.1007/s10845-013-0761-y. (2023), https://doi.org/10.34133/research.0197.
[57] A.A. Murtaza, A. Saher, M.H. Zafar, S.K.R. Moosavi, M.F. Aftab, F. Sanfilippo, [82] G. Mattera, A. Caggiano, L. Nele, Optimal data-driven control of manufacturing
Paradigm Shift for Predictive Maintenance and Condition Monitoring from processes using reinforcement learning: an application to wire arc additive
Industry 4.0 to Industry 5.0: a Systematic Review, Challenges and Case Study, manufacturing, J. Intell. Manuf. (2024), https://doi.org/10.1007/s10845-023-
Results. Eng. (2024) 102935, https://doi.org/10.1016/j.rineng.2024.102935. 02307-w.
[58] X. Yao, S.K. Moon, G. Bi, A hybrid machine learning approach for additive [83] M. Khanzadeh, P. Rao, R. Jafari-Marandi, B.K. Smith, M.A. Tschopp, L. Bian,
manufacturing design feature recommendation, Rapid. Prototyp. J. 23 (2017) Quantifying Geometric Accuracy with Unsupervised Machine Learning: using
983–997, https://doi.org/10.1108/RPJ-03-2016-0041. Self-Organizing Map on Fused Filament Fabrication Additive Manufacturing
[59] G. Williams, N.A. Meisel, T.W. Simpson, C. McComb, Design repository Parts, Journal of Manufacturing Science and Engineering, Transactions of the
effectiveness for 3D convolutional neural networks: application to additive ASME 140 (2018), https://doi.org/10.1115/1.4038598.
manufacturing, Journal of Mechanical Design 141 (2019) 111701, https://doi. [84] B. Kappes, S. Moorthy, D. Drake, H. Geerlings, A. Stebner, Machine learning to
org/10.1115/1.4044199. optimize additive manufacturing parameters for laser powder bed fusion of
[60] K. Guo, M.J. Buehler, A semi-supervised approach to architected materials design Inconel 718, in: Proceedings of the 9th International Symposium on Superalloy
using graph neural networks, Extreme Mech. Lett. 41 (2020), https://doi.org/ 718 & Derivatives: Energy, Aerospace, and Industrial Applications, Springer,
10.1016/j.eml.2020.101029. 2018, pp. 595–610, https://doi.org/10.1007/978-3-319-89480-5_39.
[61] H. Wu, J. He, H. Cheng, L. Yang, H.J. Park, J. Li, Development and analysis of [85] K. Wasmer, T. Le-Quang, B. Meylan, S.A. Shevchik, Situ Quality Monitoring in AM
machine-learning guided flash nanoprecipitation (FNP) for continuous chitosan Using Acoustic Emission: a Reinforcement Learning Approach, J. Mater. Eng.
nanoparticles production, Int. J. Biol. Macromol. 222 (2022) 1229–1237, https:// Perform. 28 (2019) 666–672, https://doi.org/10.1007/s11665-018-3690-2.
doi.org/10.1016/j.ijbiomac.2022.09.202. [86] M.H. Kunkel, A. Gebhardt, K. Mpofu, S. Kallweit, Quality assurance in metal
[62] T.A. Burge, J.R.T. Jeffers, C.W. Myant, Development of an automated mass- powder bed fusion via deep-learning-based image classification, Rapid. Prototyp.
customization pipeline for knee replacement surgery using biplanar X-Rays, J. 26 (2020) 259–266, https://doi.org/10.1108/RPJ-03-2019-0066.
Journal of Mechanical Design 144 (2022), https://doi.org/10.1115/1.4052192. [87] C.S. O’reilly, M. Elbadawi, N. Desai, S. Gaisford, A.W. Basit, M. Orlu, Machine
[63] H. Zheng, H. Mofatteh, M. Hablicsek, A. Akbarzadeh, M. Akbarzadeh, Dragonfly- learning and machine vision accelerate 3d printed orodispersible film
inspired wing design enabled by machine learning and Maxwell’s reciprocal development, Pharmaceutics. 13 (2021), https://doi.org/10.3390/
diagrams, Adv. Sci. 10 (2023), https://doi.org/10.1002/advs.202207635. pharmaceutics13122187.
[64] N. Kim, H. Lee, G. Han, M. Kang, S. Park, D.E. Kim, M. Lee, M.-J. Kim, Y. Na, [88] S. Nasiri, M.R. Khosravani, Machine learning in predicting mechanical behavior
S. Oh, S.R. Shin, H.-D. Jung, 3D-printed functional hydrogel by DNA-induced of additively manufactured parts, J. Mater. Res. Technol. 14 (2021) 1137–1153,
biomineralization for accelerated diabetic wound healing, Advanced Science 10 https://doi.org/10.1016/j.jmrt.2021.07.004.
(2023), https://doi.org/10.1002/advs.202300816. [89] J.J. Ong, B.M. Castro, S. Gaisford, P. Cabalar, A.W. Basit, G. Pérez, A. Goyanes,
[65] E. Yangue, D. Fullington, O. Smith, W. Tian, C. Liu, Diffusion generative model- Accelerating 3D printing of pharmaceutical products using machine learning, Int.
based learning for smart layer-wise monitoring of additive manufacturing, J. Pharm. X. 4 (2022), https://doi.org/10.1016/j.ijpx.2022.100120.
J. Comput. Inf. Sci. Eng. 24 (2024) 060903, https://doi.org/10.1115/1.4065092. [90] A.S. Khusheef, M. Shahbazi, R. Hashemi, Deep learning-based multi-sensor fusion
[66] M. Mohseni, S. Khodaygan, Design for additive manufacturing of topology- for process monitoring: application to fused deposition modeling, Arab. J. Sci.
optimized structures based on deep learning and transfer learning, Rapid. Eng. (2023), https://doi.org/10.1007/s13369-023-08340-4.
Prototyp. J. (2024), https://doi.org/10.1108/RPJ-02-2024-0102. [91] A.S. Raihan, H. Khosravi, T.H. Bhuiyan, I. Ahmed, An augmented surprise-guided
[67] M. Wu, V.V. Phoha, Y.B. Moon, A.K. Belman, Detecting malicious defects in 3D sequential learning framework for predicting the melt pool geometry, J. Manuf.
printing process using machine learning and image classification, in: ASME Syst. 75 (2024) 56–77, https://doi.org/10.1016/j.jmsy.2024.05.023.
International Mechanical Engineering Congress and Exposition, American Society [92] I. Rojek, J. Kopowski, J. Lewandowski, D. Mikołajewski, Use of Machine Learning
of Mechanical Engineers, 2016 V014T07A004, https://doi.org/10.1115/ to improve additive manufacturing processes, Appl. Sci. 14 (2024) 6730, https://
IMECE2016-67641. doi.org/10.3390/app14156730.
[68] F. Caiazzo, A. Caggiano, Laser direct metal deposition of 2024 al alloy: trace [93] D.R. Gunasegaram, A.B. Murphy, A. Barnard, T. DebRoy, M.J. Matthews,
geometry prediction via machine learning, Materials. (Basel) 11 (2018), https:// L. Ladani, D. Gu, Towards developing multiscale-multiphysics models and their
doi.org/10.3390/ma11030444. surrogates for digital twins of metal additive manufacturing, Addit. Manuf. 46
[69] M.G. Mohammadi, D. Mahmoud, M. Elbestawi, On the application of machine (2021), https://doi.org/10.1016/j.addma.2021.102089.
learning for defect detection in L-PBF additive manufacturing, Opt. Laser. [94] C.D. Christou, G. Tsoulfas, Role of three-dimensional printing and artificial
Technol. 143 (2021) 107338, https://doi.org/10.1016/j.optlastec.2021.107338. intelligence in the management of hepatocellular carcinoma: challenges and
[70] M.Rezapour Sarabi, M.M. Alseed, A.A. Karagoz, S. Tasoglu, Machine learning- opportunities, World J. Gastrointest. Oncol. 14 (2022) 765–793, https://doi.org/
enabled prediction of 3D-printed microneedle features, Biosensors. (Basel) 12 10.4251/wjgo.v14.i4.765.
(2022) 491, https://doi.org/10.3390/bios12070491. [95] T.M. Hamdy, How Artificial Inelegance Is Transforming Aesthetic Dentistry: a
[71] N. Jyeniskhan, A. Keutayeva, G. Kazbek, M.H. Ali, E. Shehab, Integrating machine Review, Curr. Oral Health Rep. 11 (2024) 95–104, https://doi.org/10.1007/
learning model and digital twin system for additive manufacturing, IEEe Access. s40496-024-00372-5.
11 (2023) 71113–71126, https://doi.org/10.1109/ACCESS.2023.3294486. [96] P.M. Abhilash, J. Boban, A. Ahmed, X. Luo, Digital twin-driven additive
[72] D. Chowdhury, A. Sinha, D. Das, XAI-3DP: diagnosis and understanding faults of manufacturing: advancements and future prospects. Hybrid Metal Additive
3-D printer with explainable ensemble AI, IEEe Sens. Lett. 7 (2023) 1–4, https:// Manufacturing, CRC Press, 2023, pp. 196–221, https://doi.org/10.1201/
doi.org/10.1109/LSENS.2022.3228327. 9781003406488-12.
[73] H. Bordekar, N. Cersullo, M. Brysch, J. Philipp, C. Hühne, eXplainable artificial [97] M.R. Wilkinson, B. Castro-Dominguez, C.C. Wilson, U. Martinez-Hernandez, Low-
intelligence for automatic defect detection in additively manufactured parts using cost, autonomous microscopy using deep learning and robotics: a crystal
CT scan analysis, J. Intell. Manuf. (2023), https://doi.org/10.1007/s10845-023- morphology case study, Eng. Appl. Artif. Intell. 126 (2023), https://doi.org/
02272-4. 10.1016/j.engappai.2023.106985.
[74] M. Demetgul, W. Wang, J. Fleischer, I.N. Tansel, AI-based inspection of the axes [98] M. Khanzadeh, S. Chowdhury, M. Marufuzzaman, M.A. Tschopp, L. Bian, Porosity
of machine tools, Int. J. Adv. Manuf. Technol. 130 (2024) 2329–2342, https:// prediction: supervised-learning of thermal history for direct laser deposition,
doi.org/10.1007/s00170-023-12830-y. J. Manuf. Syst. 47 (2018) 69–82, https://doi.org/10.1016/j.jmsy.2018.04.001.
[75] R. Ghimire, A. Raji, Use of Artificial Intelligence in Design, Development, [99] A. Pazienza, D. Monte, Introducing the Monitoring Equipment Mask
Additive Manufacturing, and Certification of Multifunctional Composites for Environment, Sensors 22 (2022), https://doi.org/10.3390/s22176365.
Aircraft, Drones, and Spacecraft, Appl. Sci. (Switzerland) 14 (2024), https://doi. [100] M.A.D. Silva, B. Amaro Junior, R.R.B. Medeiros, P.R. Pinheiro,
org/10.3390/app14031187. A neuroevolutionary model to estimate the tensile strength of manufactured parts
[76] S. Stoyanov, C. Bailey, Machine learning for additive manufacturing of made by 3D printing, Algorithms. 15 (2022), https://doi.org/10.3390/
electronics, in: 2017 40th International Spring Seminar on Electronics a15080263.
Technology (ISSE), IEEE, 2017, pp. 1–6, https://doi.org/10.1109/ [101] N. Pesce, M. Fortunato, A. Tamburrano, 3-D printed graphene-based
ISSE.2017.8000936. piezoresistive foam mat for pressure detection through electrical resistance
[77] J.M. Gardner, K.A. Hunt, A.B. Ebel, E.S. Rose, S.C. Zylich, B.D. Jensen, K.E. Wise, tomography and machine learning classification techniques, IEEe Sens. Lett. 7
E.J. Siochi, G. Sauti, Machines as Craftsmen: localized Parameter Setting (2023), https://doi.org/10.1109/LSENS.2023.3307077.
Optimization for Fused Filament Fabrication 3D Printing, Adv. Mater. Technol. 4 [102] A. Mishra, V.S. Jatti, Neurosymbolic artificial intelligence (NSAI) based algorithm
(2019), https://doi.org/10.1002/admt.201800653. for predicting the impact strength of additive manufactured polylactic acid (PLA)
[78] C. Silbernagel, A. Aremu, I. Ashcroft, Using machine learning to aid in the specimens, Eng. Res. Express 5 (2023), https://doi.org/10.1088/2631-8695/
parameter optimisation process for metal-based additive manufacturing, Rapid. ace610.
Prototyp. J. 26 (2020) 625–637, https://doi.org/10.1108/RPJ-08-2019-0213. [103] H. Izadgoshasb, A. Kandiri, P. Shakor, V. Laghi, G. Gasparini, Predicting
[79] P. Oehlmann, P. Osswald, J.C. Blanco, M. Friedrich, D. Rietzel, G. Witt, Modeling compressive strength of 3D printed mortar in structural members using machine
Fused Filament Fabrication using Artificial Neural Networks, Production learning, Appl. Sci. 11 (2021) 10826, https://doi.org/10.3390/app112210826.
Engineering 15 (2021) 467–478, https://doi.org/10.1007/s11740-021-01020-y. [104] Z. Zheng, S. Qiu, X. Yue, J. Wang, J. Hou, Detecting irradiation defects in
[80] S. Rahmani Dabbagh, O. Ozcan, S. Tasoglu, Machine learning-enabled materials: a machine learning approach to analyze helium bubble images, J. Nucl.
optimization of extrusion-based 3D printing, Methods 206 (2022) 27–40, https:// Mater. 596 (2024), https://doi.org/10.1016/j.jnucmat.2024.155117.
doi.org/10.1016/j.ymeth.2022.08.002. [105] J. Lizarazu, E. Harirchian, U.A. Shaik, M. Shareef, A. Antoni-Zdziobek, T. Lahmer,
Application of machine learning-based algorithms to predict the stress-strain

26
S. Inayathullah and R. Buddala Results in Engineering 25 (2025) 103676

curves of additively manufactured mild steel out of its microstructural [116] M. Danish, M.K. Gupta, S.A. Irfan, S.M. Ghazali, M.F. Rathore, G.M. Krolczyk,
characteristics, Results. Eng. 20 (2023) 101587, https://doi.org/10.1016/j. A. Alsaady, Machine learning models for prediction and classification of tool wear
rineng.2023.101587. in sustainable milling of additively manufactured 316 stainless steel, Results. Eng.
[106] Z. Li, Z. Zhang, J. Shi, D. Wu, Prediction of surface roughness in extrusion-based 22 (2024) 102015.
additive manufacturing with machine learning, Robot. Comput. Integr. Manuf. 57 [117] M. Khanzadeh, P. Rao, R. Jafari-Marandi, B.K. Smith, M.A. Tschopp, L. Bian,
(2019) 488–495, https://doi.org/10.1016/j.rcim.2019.01.004. Quantifying geometric accuracy with unsupervised machine learning: using self-
[107] R. Neelam, S.A. Kulkarni, H.S. Bharath, S. Powar, M. Doddamani, Mechanical organizing map on fused filament fabrication additive manufacturing parts,
response of additively manufactured foam: a machine learning approach, Results. J. Manuf. Sci. Eng. 140 (2018) 031011, https://doi.org/10.1115/1.4038598.
Eng. 16 (2022) 100801, https://doi.org/10.1016/j.rineng.2022.100801. [118] J.J. Ong, B.M. Castro, S. Gaisford, P. Cabalar, A.W. Basit, G. Pérez, A. Goyanes,
[108] T. Sahar, M. Rauf, A. Murtaza, L.A. Khan, H. Ayub, S.M. Jameel, I.U. Ahad, Accelerating 3D printing of pharmaceutical products using machine learning, Int.
Anomaly detection in laser powder bed fusion using machine learning: a review, J. Pharm. X. 4 (2022) 100120, https://doi.org/10.1016/j.ijpx.2022.100120.
Results. Eng. 17 (2023) 100803, https://doi.org/10.1016/j.rineng.2022.100803. [119] D. Beiderbeck, H. Krüger, T. Minshall, The future of additive manufacturing in
[109] K. Garanger, T. Khamvilai, E. Feron, 3D printing of a leaf spring: a demonstration sports. 21st Century Sports: How Technologies Will Change Sports in the Digital
of closed-loop control in additive manufacturing, in: 2018 IEEE Conference on Age, 2020, pp. 111–132, https://doi.org/10.1007/978-3-030-50801-2_7.
Control Technology and Applications (CCTA), IEEE, 2018, pp. 465–470, https:// [120] X. Fu, Y. Lin, X.-J. Yue, Xun Ma, B. Hur, X.-Z. Yue, A review of additive
doi.org/10.1109/CCTA.2018.8511509. manufacturing (3D printing) in aerospace: technology, materials, applications,
[110] T. Mukherjee, T. DebRoy, A digital twin for rapid qualification of 3D printed and challenges, in: Mobile Wireless Middleware, Operating Systems and
metallic components, Appl. Mater. Today 14 (2019) 59–65, https://doi.org/ Applications: 10th International Conference on Mobile Wireless Middleware,
10.1016/j.apmt.2018.11.003. Operating Systems and Applications (MOBILWARE 2021), Springer, 2022,
[111] D. Grierson, A.E.W. Rennie, S.D. Quayle, Machine learning for additive pp. 73–98, https://doi.org/10.1007/978-3-030-98671-1_6.
manufacturing, Encyclopedia (Basel,. 2021) 1 (2021) 576–588, https://doi.org/ [121] A.M. Nayeem, M.M.N. Hossain, Usage of additive manufacturing in the
10.3390/encyclopedia1030048. automotive industry: a review, Bangladesh J. Multidiscip. Sci. Res. 8 (2023) 9–20,
[112] F.S.H.B. Freeman, L. Chechik, I. Todd, Beat the machine (learning): metal https://doi.org/10.46281/bjmsr.v8i1.2135.
additive manufacturing and closed loop control, Phys. Educ. 55 (2020) 055012, [122] C. Belvedere, A. MacLeod, A. Leardini, A. Grassi, G.D.A.L. FABBRO, S. Zaffagnini,
https://doi.org/10.1088/1361-6552/ab9957. H.S. GILL, 3D medical imaging analysis, patient-specific instrumentation and
[113] O. Mypati, A. Mukherjee, D. Mishra, S.K. Pal, P.P. Chakrabarti, A. Pal, A critical individualized implant design, with additive manufacturing creates a new
review on applications of artificial intelligence in manufacturing, Artif. Intell. personalized high tibial osteotomy treatment option, J. Mech. Med. Biol. 23
Rev. 56 (2023) 661–768, https://doi.org/10.1007/s10462-023-10535-y. (2023) 2340041, https://doi.org/10.1142/S0219519423400419.
[114] S. Nasiri, M.R. Khosravani, Machine learning in predicting mechanical behavior [123] D.R. Gunasegaram, A.B. Murphy, M.J. Matthews, T. DebRoy, The case for digital
of additively manufactured parts, J. Mater. Res. Technol. 14 (2021) 1137–1153, twins in metal additive manufacturing, Journal of Physics: Materials 4 (2021)
https://doi.org/10.1016/j.jmrt.2021.07.004. 040401, https://doi.org/10.1088/2515-7639/ac09fb.
[115] S. Yadav, S. Shukla, Analysis of k-fold cross-validation over hold-out validation on [124] A. Busachi, J. Erkoyuncu, P. Colegrove, F. Martina, C. Watts, R. Drake, A review
colossal datasets for quality classification, in: 2016 IEEE 6th International of Additive Manufacturing technology and Cost Estimation techniques for the
Conference on Advanced Computing (IACC), IEEE, 2016, pp. 78–83, https://doi. defence sector, CIRP. J. Manuf. Sci. Technol. 19 (2017) 117–128, https://doi.org/
org/10.1109/IACC.2016.25. 10.1016/j.cirpj.2017.07.001.

27