PRESENTATION

TOPIC
MF F318:
Machine Learning in Metal Additive COURSE Non Traditional
Manufacturing Manufacturing Processes

Student Name Atharva Patil Adhik Agarwal

Student ID 2022ABPS1119P 2022ABPS1364P

Contribution: Research 50% 50%

Contribution: Slide Preparation 60% 40%

 Introduction & BACKGROUND
Machine Learning (ML) ML + AM = Smart, self-learning manufacturing systems Additive Manufacturing (AM)

Machine Learning (ML) Moves towards Industry 4.0 goals with data-driven ☑ Metal Additive Manufacturing
enables AM to evolve from decision-making. (AM) fabricates components
trial-and-error to predictive layer-by-layer using
and intelligent manufacturing: techniques like:
• Supervised, Unsupervised, • Powder Bed Fusion,
Semi-supervised, and Directed Energy
Reinforcement Learning Deposition, Binder Jetting,
help learn from massive etc.
process data. ☑ AM offers high design
• Deep Learning models (e.g., freedom, material efficiency,
Neural Networks, and rapid prototyping, but is
Transformers) automate challenged by:
defect detection, parameter • Inconsistent part quality
tuning, and process • Lack of real-time control
optimization. • High post-processing
demands
State of the Art
MA ni stfiineb
LM fo yrammmuS
 Challenges Faced by ML‑Based Techniques in AM Processes
1. Data Availability and Quality

☐ ML performance depends heavily on large, high-quality datasets, which are scarce in AM.
☐ AM data is often inconsistent and scattered across machines, materials, and conditions.
☐ Experimental data collection is costly and time-consuming, limiting access to training data.

2. Model Complexity and Interpretability

☐ Due to AM’s complex multi-physics nature, ML models become highly complex and non-intuitive.
☐ Models like deep neural networks are accurate but act as black boxes, difficult to explain or trust.
☐ Lack of explainability reduces adoption in high-stakes industries (e.g., aerospace, biomedical)

3. Real-Time Data Acquisition and Processing

☐ AM systems generate large volumes of sensor data during fabrication.

☐ Processing this data in real-time requires high-speed algorithms and integration with hardware.
☐ Real-time adaptive control using ML is still limited due to system constraints and latency.
 Challenges Faced by ML‑Based Techniques in AM Processes
4. Data Privacy and Security

☐ AM workflows contain sensitive design and process data, especially in aerospace and defense.
☐ Cloud-based ML solutions risk unauthorized access and IP theft.
☐ Secure data sharing and storage frameworks are needed for industrial use.

5. Generalization and Transferability

☐ ML models trained on specific machines or materials often fail when transferred to others.
☐ Differences in equipment, sensors, and settings lead to low model robustness.
☐ Transfer learning and domain adaptation are underutilized but essential.

6. Lack of Standardized Benchmarks

☐ There is no unified benchmark to evaluate ML models in AM.

☐ Researchers use different datasets, metrics, and tasks, making comparisons hard.
☐ Standardizing evaluation protocols is crucial for reproducibility and industry trust.
ML in AM Process (Process, Mechanism)
1. Design Optimization via ML 2. Material Design
☐ Predicts part behavior under different loads and materials. ☐ Identifies optimal alloy compositions for AM.
☐ Suggests geometry improvements based on historical ☐ Predicts mechanical properties from composition +
design-performance data. process parameters.
☐ ML supports generative design for complex, non-intuitive ☐ Accelerates discovery of lightweight, durable, corrosion-
geometries. resistant materials.

Workflow of ML in material design (Physics-informed ML in the Multi-scale learning approach in AM (A hierarchical multi-
fatigue design problem) scale ML framework for piezoelectric material design)
ML in AM Process (Process, Mechanism)
3. Topology Design

☐ ML assists in finding lightweight, strong internal structures.

☐ Enhances structural performance while reducing material use.
☐ Combines CNNs + optimization to generate novel AM-suited topologies.

Use of ML in topology
design/optimization
ML in AM Design Stage (Mechanism & Methods)
1. Process Parameter Optimization
Use of ML for
☐ ML predicts optimal laser power, speed, layer process
thickness to improve part quality. parameter
☐ Improves consistency in density, dimensional optimization in
accuracy, and surface roughness. AM

2. Powder Spreading Characterization

☐ Analyzes particle flow and layer thickness using vision-based ML.
☐ Ensures uniformity in powder layers for better fusion and part accuracy.

Powder spreading
mechanism
ML in AM Design Stage (Mechanism & Methods)
3. In-Process Defect Monitoring
☑ Detects real-time defects like porosity, delamination, cracking using:
• Acoustic-based ML • Optical sensors
• Thermal imaging • Vibration signals

Acoustic-based monitoring in AM

Optic-based monitoring in AM Thermal-based monitoring in AM Vibration-based monitoring in AM

ML in AM Production (Methods & Planning)
1. Planning
☐ ML optimizes build orientation &
supports.
☐ Predicts build time and resource usage.

2. Quality Control

☐ Compares build output with CAD models.

☐ Ensures geometric fidelity and
tolerances.

3. Data Security
☐ Detects cyber-intrusions and data leaks.
☐ Encrypts design files during cloud
sharing.

ML-based quality assurance framework

 Analytical/theoretical/ mathematical models
Supervised Learning

LR

SVM

RF RNN,GRU,LSTM CNN
 Analytical/theoretical/ mathematical models
DIAGRAMS
unSupervised
Learning
Models
Hybrid
 ML MODAL APPLICATION IN AM PROCESSES
1. Material Extrusion (MEX) 2. Direct Energy Deposition (DED) 3.Material Jetting (MJ)

4.Powder Bed Fusion (PBF) 5. Sheet lamination (SL) 6. Vat Photopolymerization (VP)
ML MODAL APPLICATION IN AM PROCESSES

ML approaches in design for AM ML approaches in the AM process

 Case Study – Lightweight Design in a Remote-Controlled Car Using ML
Objective:
To reduce weight of critical RC car components (bumper,
wheel, suspension arm, driveshaft) using machine learning-
driven AM feature recommendations, while retaining
structural performance.

1. Initial Traditional Design Stage (Pre-ML)

☐ Components: Bumper, Wheel, Suspension Arm,

Driveshaft.
☐ Nature: Traditional design geometry, often solid and
bulky — not ideal for Additive Manufacturing (AM).
☐ Limitation:
• Heavyweight.
• Low design complexity.
• Not optimized for AM constraints like overhangs,
material usage, and print time.
Case Study – Lightweight Design in a Remote-Controlled Car Using ML
2. ML-Based AM Design Feature Recommendation Workflow

This is the central mechanism where ML enters the picture and transforms how design recommendations are generated.
➤ Step-by-Step Breakdown:
Step 1: Data Preparation – Encoding Step 3: Learning from the Past – Supervised Classification
• Multi-Category Encoding: Converts design attributes of • SVM (Support Vector Machine) Classifier:
target components into numeric categorical values. • Trained on the clustered dataset with historical examples.
• Purpose: Make them readable for machine learning • Learns the mapping between component types and
algorithms. successful AM features.
• Example: A wheel may be encoded based on shape, • Input: Component characteristics.
function, symmetry, etc. • Output: Best-suited AM feature category.

Step 2: Feature Similarity – Clustering Step 4: Prediction & Recommendation

• Hierarchical Clustering & Dendrogram Generation: • Final sub-cluster identified by SVM contains design features
• Groups design features from existing successful recommended for AM adaptation of the original component.
industrial AM applications based on structural or • These features include:
functional similarity. • Lattice infill for bumpers.
• Unsupervised ML step: Discovers natural groupings of • Ribbed or skeletal structures for arms.
AM features (like lattices, ribs, trusses, hollow sections). • Hollowed or spoked structures for wheels.
 Case Study – Lightweight Design in a Remote-Controlled Car Using ML
3. Final Design Implementation (Post-ML)

Transformation: Original designs replaced with:

• Lightweight, complex geometries.
• AM-friendly shapes (less support, optimized
strength-to-weight ratio).

Examples:
• Lattice bumper.
• Spoked wheel.
• Trussed suspension arm.
 Solutions to Key Challenges in ML for AM
1. Data Availability & Quality

☐ Develop benchmark datasets via collaboration between academia & industry.

☐ Use simulation data for model pre-training.
☐ Employ data augmentation techniques (bootstrapping, synthetic image generation).

2. Model Complexity & Interpretability

☐ Shift toward Explainable AI (XAI).

☐ Use interpretable models like decision trees and SHAP values.
☐ Integrate Physics-Informed ML (PIML) to retain domain logic.

3. Real-Time Data Acquisition and Processing

☐ Combine IoT-based sensors with lightweight deep learning models.

☐ Use edge computing to reduce latency.
☐ Implement CNN+LSTM for sequence-based predictions.
 Solutions to Key Challenges in ML for AM
4. Data Privacy and Security

☐ Use ML for intrusion detection and anomaly tracking.

☐ Implement biometric & behavioral ML-based access control.
☐ Encrypt sensitive production data.

5. Generalization and Transferability

☐ Build cross-domain ML models using federated learning.

☐ Train models on multi-process, multi-material datasets.
☐ Use transfer learning to apply trained models to new AM systems.

6. Lack of Standardized Benchmarks

☐ Establish open-source validation protocols.

☐ Adopt unified metrics (e.g., accuracy, RMSE, MAPE) across research groups.
☐ Propose universal dataset repositories and cloud model evaluations.
 Future Perspectives
1. Generative Design & Topology Optimization 2. Real-Time Monitoring & Predictive Maintenance
☐ ML-driven generative design will create complex, ☐ ML will enable preemptive fault detection and
high-performance geometries. adaptive control, minimizing downtime.
☐ Enhances structural efficiency while reducing
material usage and time.

3. Microstructural Characterization 4. Microstructure-to-Property Linkage

☐ ML automates microstructure analysis from ☐ ML correlates grain size, porosity, phase
imaging data. distribution with properties like strength and
☐ Predicts material behavior based on internal fatigue.
structure, improving quality control. ☐ Enables tailored material development for
specific applications.
5. New Material Discovery 6. Experiments, Simulation, and ML in Parallel
☐ ML accelerates discovery of novel alloys and ☐ Integrating all three enhances process efficiency
optimal compositions for AM. and design decision-making.
☐ Extracts insights from vast experimental and ☐ Promotes fast feedback loops for real-time
simulation datasets. optimization.
Conclusion
☐ Machine Learning (ML) is transforming Metal Additive Manufacturing (AM) from a trial-and-error
process into a data-driven, intelligent production system.
☐ It enables:
• Smarter design through generative and topology optimization.
• More efficient processing via real-time monitoring and parameter tuning.
• Improved material performance through microstructure-property linkage and predictive modeling.
☐ ML also accelerates:
• New material development
• Automated quality control
• Informed decision-making across the AM workflow
☐ As ML models become more generalizable, interpretable, and integrated with AM systems, the
industry is poised for a leap in efficiency, reliability, and innovation.
☐ The future of AM is not just additive — it's adaptive, intelligent, and AI-driven.