See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/383027298

Retrieval-Augmented Generation in Engineering Design

Preprint · August 2024

DOI: 10.13140/RG.2.2.18595.26406

CITATION READS

1 1,007

1 author:

Debi Prasad Ghosh

Larsen & Toubro (L&T) Construction
110 PUBLICATIONS 101 CITATIONS

SEE PROFILE

All content following this page was uploaded by Debi Prasad Ghosh on 11 August 2024.

The user has requested enhancement of the downloaded file.

Manuscript as of August 09, 2024

Retrieval-Augmented Generation in Engineering Design

Debi Prasad Ghosh, [email protected]

Design Automation Team

Engineering Design & Research Center, Larsen & Toubro Construction (M&M) Kolkata, India

Abstract: This paper explores the application of Retrieval-

Augmented Generation (RAG) in engineering design, examining its
potential to revolutionize the design process through advanced
computational techniques and artificial intelligence. We
investigate the core components of the RAG framework—Retrieve,
Analyze, and Generate—and their contribution to enhanced
efficiency, improved accuracy, and increased innovation in
engineering design. The study discusses advanced techniques
that augment RAG capabilities, including Physics-Informed
Geometry-Aware Neural Operators (PI-GANO), Geometric
Operators (GOs), and Multi-Fidelity Cross-Validation (MFCV).
These methods significantly improve design simulation accuracy,
reduce surrogate model overfitting, and optimize computational
resource utilization. We explore emerging technologies facilitating
RAG implementation, such as Azure Intelligent Document, Azure
OpenAI, LlamaIndex, and LangChain, demonstrating their role in
automating data extraction, enhancing natural language
processing, and enabling efficient information retrieval in
engineering contexts. The paper addresses challenges in RAG
adoption, including data quality, computational requirements, and
the need for interpretability in AI-driven design decisions.
This comprehensive exploration of RAG in engineering design
contributes to the intersection of artificial intelligence and
engineering, offering insights for researchers and practitioners. As
RAG evolves, it promises to reshape engineering design, enabling
faster, more accurate, and more creative solutions to complex
challenges.
1. Introduction
Engineering design has always been at the forefront of technological
innovation, constantly evolving to meet the challenges of an increasingly
complex world. As the demands for more efficient, sustainable, and
innovative solutions grow, so does the need for advanced methodologies
that can leverage the vast amounts of available data and knowledge. In
recent years, the integration of artificial intelligence (AI) and advanced
computational techniques has opened new frontiers in design
methodologies. Among these emerging approaches, Retrieval-Augmented
Generation (RAG) stands out as a particularly promising framework for
enhancing the design process.
RAG combines the power of large-scale information retrieval with
sophisticated analysis and generative capabilities. This synergy allows
engineers to leverage vast amounts of existing knowledge while
simultaneously pushing the boundaries of innovation. By automating and
optimizing various aspects of the design process, RAG has the potential to
significantly reduce development time, improve design quality, and foster
creativity in ways previously unattainable.
The core components of the RAG framework—Retrieve, Analyze, and
Generate—work in harmony to create a powerful tool for engineering design:
1. The Retrieve stage gathers relevant data from diverse sources,
including historical design data, simulation results, and existing
literature.
2. The Analyze phase evaluates this data to extract insights and identify
patterns that inform the design process.
3. The Generate stage produces innovative design solutions based on
the analysis of the retrieved data.
This paper aims to provide a comprehensive overview of RAG in the context
of engineering design. We will explore the theoretical foundations of the RAG
framework, examine its practical applications across different engineering
domains, and discuss the tools and technologies that enable its
implementation. Furthermore, we will analyze the benefits and challenges
associated with RAG adoption and consider future directions for research
and development in this rapidly evolving field.
By investigating emerging technologies such as Azure Intelligent Document,
Azure OpenAI, LlamaIndex, and LangChain, we will demonstrate how these
tools facilitate the implementation of RAG approaches in real-world
engineering scenarios. Through a comprehensive analysis of current
methodologies and case studies, we will showcase the significant impact of
RAG on reducing project costs and improving design outcomes across
various engineering domains, including structural engineering, mechanical
design, aerospace engineering, and robotics.
As we delve into the intricacies of RAG and its applications, this paper seeks
to illuminate the transformative potential of this approach in revolutionizing
engineering design practices. By the end of this exploration, readers will gain
a deep understanding of how RAG is shaping the future of engineering
innovation and paving the way for more efficient, accurate, and creative
design solutions.
1.1 Literature Survey: Retrieval-Augmented Generation in Engineering Design
Retrieval-Augmented Generation (RAG) has emerged as a powerful
approach in engineering design, combining the strengths of information
retrieval, data analysis, and generative techniques. This literature survey
provides an overview of recent advancements in RAG and related
technologies that are shaping the future of engineering design.
1.1.2 Advanced Techniques in RAG
1.1.2.1 Physics-Informed Neural Operators
Zhao et al. [1] introduced the Physics-Informed Geometry-Aware Neural
Operator (PI-GANO), a method for solving parametric Partial Differential
Equations (PDEs) with varying domain geometries. PI-GANO combines
physics-informed learning with a geometry encoder, offering improved
accuracy and efficiency without requiring large datasets.
Building on this, Zhao et al. [32] developed the Diffeomorphism Neural
Operator (DNO), which maps different shapes onto a standard domain and
then learns a neural operator to solve PDEs on this standard domain. This
approach improves generalization to different shapes and parameters while
reducing computational costs.
1.1.2.2 Geometric Operators and Surrogate Models
Khan et al. [3] introduced Physics-Informed Geometric Operators (GOs) to
enhance the performance of surrogate, dimension reduction, and
generative models in engineering design. GOs extract high-level geometric
information and physics from shapes, improving model performance and
reducing overfitting.
Renganathan and Carlson [4] proposed Multi-fidelity Cross-validation
(MFCV), a method for improving surrogate models by combining multiple
fidelity models into a single surrogate model. This approach offers more
efficient use of computational resources and improved accuracy.
1.1.2.3 Bayesian Optimization and Uncertainty Quantification
Ahmadianshalchi et al. [86] introduced PAC-MOO, a Bayesian optimization
method for handling constrained multi-objective optimization problems.
This approach uses surrogate models for objectives and constraints to
efficiently explore the design space.
Nemani et al. [80] provided a comprehensive overview of uncertainty
quantification (UQ) for machine learning models, focusing on neural
networks. The paper discusses various UQ techniques and their
applications in engineering design and health prognostics.
1.1.3. AI-Driven Design and Optimization
1.1.3.1 Generative Models for Engineering Design
Chong et al. [2] proposed a method to improve the practicality of using text-
to-image models for engineering design by incorporating CAD images into
model prompts. This approach enhances design feasibility while
maintaining creativity.
Fan et al. [66] introduced SA-ALAE, a deep generative model for creating
realistic engineering designs. The model incorporates self-attention to
capture long-range dependencies in engineering designs, enabling the
generation of realistic and diverse engineering blueprints.
1.1.3.2 Optimization Algorithms
Lou et al. [24] introduced the Competitive Game Optimizer (CGO), a new
optimization algorithm inspired by competitive game dynamics. CGO
incorporates exploration and exploitation phases and has demonstrated
effectiveness in UAV path planning and other engineering design problems.
Jiang and Luo [30] developed AutoTRIZ, a tool that combines the Theory of
Inventive Problem Solving (TRIZ) with large language models to automate the
design ideation process. AutoTRIZ offers a more accessible and efficient
approach to innovation in engineering design.
1.1.3.3 Machine Learning in Design Processes
Wong et al. [7] introduced Prompt Evolution Design Optimization (PEDO), a
method that uses natural language prompts to generate 3D designs and then
evaluates and improves these designs using physics-based simulations and
a vision-language model.
Park and Kang [9] proposed a Bayesian graph neural network (GNN) model
for predicting engineering performance directly from mesh representations
of CAD models. This approach improves accuracy in predicting engineering
performance compared to other 3D representation methods.
1.1.4 Data-Driven Approaches and Datasets
1.1.4.1 Engineering Design Datasets
Hong et al. [8] introduced the DeepJEB dataset, a large-scale collection of
3D jet engine bracket designs and their corresponding structural analysis
data. This dataset supports improved surrogate model performance and is
compatible with complex models like graph neural networks.
Cobb et al. [75] created AircraftVerse, a massive dataset containing diverse
aerial vehicle designs and their corresponding performance metrics. This
dataset serves as a valuable resource for researchers developing AI-driven
aircraft design tools and optimization techniques.
1.1.4.2 Data-Driven Design Methodologies
Yang et al. [96] proposed a framework for Data-Driven Intelligent
Computational Design (DICD), utilizing deep learning to extract design
features from historical data and generate novel designs. The framework
provides a systematic roadmap for DICD implementation, including dataset
building, feature engineering, and design optimization.
Picard et al. [77] presented a systematic approach to creating synthetic
datasets for engineering design applications. The research provides
guidelines for generating, annotating, and validating synthetic datasets,
addressing the critical challenge of data availability in AI-driven engineering
design.
1.1.5 Human-AI Collaboration in Engineering Design
1.1.5.1 Evaluating AI Performance in Design Tasks
Ege et al. [23] compared ChatGPT 4.0 to human engineering students in a
design challenge. The study found that ChatGPT demonstrated strong
concept generation abilities but struggled with problem-solving and
decision-making. The research provides recommendations for effective
human-AI collaboration in design.
Doris et al. [25] introduced DesignQA, a benchmark to evaluate how well
large language models understand and apply engineering requirements
from technical documents. The study establishes a foundation for improving
AI-assisted engineering design processes.
1.1.5.2 Collaborative Design Tools
Zhu et al. [16] investigated the effectiveness of human-AI collaboration in
feature engineering. The study developed a prototype tool to provide AI and
human-generated feature suggestions to data scientists and conducted a
user study to understand how users interact with and perceive AI and
human-generated features.
Gmeiner et al. [91] investigated how engineers learn to use AI-based design
tools, focusing on understanding the challenges designers face when
collaborating with AI tools in a design process. The study highlights the need
for improved AI tool design to facilitate effective human-AI collaboration.
1.1.6 Applications in Specific Engineering Domains
1.1.6.1 Aerospace Engineering
Yu et al. [26] introduced PFN4sBO, a method for optimizing complex
engineering problems using Bayesian Optimization. The approach uses pre-
trained transformers to quickly and accurately find optimal solutions, with
potential applications in various engineering fields, including aerospace.
1.1.6.2 Mechanical Engineering
Nobari et al. [10] developed LInK, a method for designing mechanisms that
combines deep learning (contrastive learning) with traditional optimization
techniques. LInK offers significantly faster and more accurate optimization
than previous methods, especially for complex problems.
1.1.6.3 Robotics and Control Systems
Tong et al. [5] focused on determining the initial shape of flexible rods for
desired deformation under gravity. The research combines physics-based
modeling with machine learning techniques, with potential applications in
soft robotics and animation.
Kevian et al. [27] evaluated the capabilities of large language models in
solving control engineering problems. The study developed a benchmark
dataset (ControlBench) and compared various LLMs on undergraduate-level
control problems.
1.1.7 Future Directions and Challenges
As the field of RAG in engineering design continues to evolve, several
challenges and opportunities emerge:
1. Data Quality and Availability: Ensuring comprehensive and accurate
datasets across all relevant engineering domains remains a challenge,
as highlighted by Picard et al. [77].
2. Computational Resources: Advanced RAG techniques often require
significant computational power, which may be a limitation for smaller
organizations or resource-constrained projects.
3. Integration with Existing Workflows: Incorporating RAG methodologies
into established engineering processes may require substantial
changes to existing workflows and training for personnel.
4. Interpretability and Explainability: As RAG systems become more
complex, ensuring the interpretability and explainability of design
decisions becomes increasingly important, particularly in safety-
critical applications.
5. Human-AI Collaboration: Improving the interface between AI systems
and human engineers to leverage the strengths of both is an ongoing
area of research, as demonstrated by studies like Gmeiner et al. [91].
6. Ethical Considerations: As AI plays an increasingly significant role in
engineering design, addressing ethical considerations and ensuring
responsible AI development becomes crucial, as discussed by
Constantinides et al. [65].
1.1.8 Conclusion: Retrieval-Augmented Generation in Engineering Design
Retrieval-Augmented Generation is revolutionizing engineering design by
enabling more efficient, accurate, and innovative design processes. The
integration of physics-informed methods, advanced optimization
techniques, and large-scale datasets is pushing the boundaries of what's
possible in engineering design. As the field continues to evolve, addressing
challenges related to data quality, computational resources, and human-AI
collaboration will be crucial for realizing the full potential of RAG in
engineering design.
1.2 Literature Review: Emerging Technologies in RAG for Engineering Design
1.2.1. Introduction: Emerging Technologies
Retrieval-Augmented Generation (RAG) has emerged as a powerful
approach in engineering design, combining the strengths of information
retrieval, data analysis, and generative techniques. This literature review
focuses on four key technologies that are facilitating the implementation of
RAG approaches in real-world engineering scenarios: Azure Intelligent
Document [138], Azure OpenAI [139], LlamaIndex [140], and LangChain
[141].
1.2.2. Azure Intelligent Document
Azure Intelligent Document is a cloud-based service that automates the
extraction of relevant data from various document formats, streamlining the
information retrieval process crucial for RAG systems.
1.2.2.1 Key Features and Capabilities
- Optical Character Recognition (OCR): Azure Intelligent Document
incorporates advanced OCR capabilities to convert images and
scanned documents into machine-readable text. This is particularly
valuable in engineering contexts where historical design documents
or hand-drawn sketches need to be digitized and analyzed.
- Form Recognition: The service can automatically identify and extract
key-value pairs, tables, and structured data from forms and
documents. This capability is essential for processing engineering
specifications, bills of materials, and other structured documents
commonly used in design processes.
- Custom Model Training: Azure Intelligent Document allows users to
train custom models on domain-specific documents, enabling more
accurate extraction of engineering-specific information.
1.2.2.2 Applications in Engineering Design
Knaster et al. [134] highlight the importance of efficient document
processing in large-scale engineering projects like the International Fusion
Materials Irradiation Facility (IFMIF). While not explicitly mentioning Azure
Intelligent Document, their work underscores the need for tools that can
efficiently process and extract information from complex engineering
reports and design documents.
Harish and Prasad [136] present a method to convert 2D engineering
drawings into 3D models using computer vision techniques. Although their
work focuses on OpenCV, it demonstrates the potential for services like
Azure Intelligent Document to automate the digitization and analysis of
engineering drawings, a crucial step in implementing RAG systems for legacy
design data.
1.2.3. Azure OpenAI
Azure OpenAI provides access to powerful language models for natural
language processing and design ideation, enabling sophisticated text
generation and completion tasks within RAG systems.
1.2.3.1 Key Features and Capabilities
- Text Generation and Completion: Azure OpenAI's language models
can generate coherent and contextually relevant text, which is
valuable for creating design descriptions, specifications, and even
code snippets.
- Semantic Similarity Analysis: The service can perform semantic
similarity comparisons, enabling the identification of related design
concepts or principles across different documents or projects.
- Fine-tuning Capabilities: Azure OpenAI allows for fine-tuning of
models on domain-specific data, enabling more accurate and
relevant outputs for engineering applications.
1.2.3.2 Applications in Engineering Design
Ege et al. [23] compared ChatGPT 4.0 to human engineering students in a
design challenge. While their study focused on ChatGPT specifically, it
demonstrates the potential of large language models like those available
through Azure OpenAI in engineering design tasks. The research highlights
the strengths of AI models in concept generation and their current
limitations in problem-solving and decision-making.
Wong et al. [7] introduced Prompt Evolution Design Optimization (PEDO), a
method that uses natural language prompts to generate 3D designs. While
not specifically using Azure OpenAI, their work illustrates how large
language models can be integrated into the design process, suggesting
potential applications for Azure OpenAI in engineering design ideation and
optimization.
1.2.4. LlamaIndex
LlamaIndex is a data framework for large language models that enables
efficient retrieval of relevant information from large datasets, making it
particularly useful for implementing RAG systems in engineering design
contexts.
1.2.4.1 Key Features and Capabilities
- Advanced Indexing Techniques: LlamaIndex employs sophisticated
indexing methods to enable quick and accurate information retrieval
from large, unstructured datasets.
- Query Processing: The framework offers customizable query
processing capabilities, allowing users to tailor information retrieval
to specific engineering domains.
- Integration with Various Data Sources: LlamaIndex can seamlessly
integrate with multiple data sources and formats commonly used in
engineering design, including text documents, CAD files, and
simulation results.
1.2.4.2 Applications in Engineering Design
While LlamaIndex is a relatively new tool, its potential applications in
engineering design are significant. Siddharth and Luo [48] investigated the
underlying structure of engineering design knowledge as expressed in patent
documents. Their work in creating knowledge graphs from patent text aligns
well with LlamaIndex's capabilities in processing and indexing large volumes
of technical documentation.
Hong et al. [8] introduced the DeepJEB dataset, a large-scale collection of
3D jet engine bracket designs and their corresponding structural analysis
data. While not directly using LlamaIndex, their work highlights the need for
efficient data management and retrieval systems in engineering design, a
need that LlamaIndex is well-positioned to address.
1.2.5. LangChain
LangChain is a framework for developing applications powered by language
models, facilitating the creation of interactive design tools and complex RAG
systems.
1.2.5.1 Key Features and Capabilities
- Modular Architecture: LangChain provides a modular architecture for
combining different language model capabilities, enabling the
creation of sophisticated RAG systems tailored to engineering design
needs.
- Context and Memory Management: The framework offers tools for
managing context and memory in language model interactions,
crucial for maintaining coherence in complex design tasks that may
span multiple sessions or involve iterative refinement.
- Integration with External Tools: LangChain can integrate with external
tools and data sources, enhancing the versatility of RAG applications
in engineering contexts.
1.2.5.2 Applications in Engineering Design
Göpfert et al. [73] explored the potential of large language models to
revolutionize the engineering design process. While not specifically
mentioning LangChain, their work aligns with the framework's capabilities in
integrating language models into complex workflows. They propose using
LLMs to automate creative and reasoning aspects of the design process,
which could be implemented using LangChain's modular architecture.
Rios et al. [68] investigated the integration of large language models and text-
to-3D models for engineering design optimization. Their work demonstrates
the potential for frameworks like LangChain to bridge the gap between
natural language processing and geometric modeling in engineering design.
1.2.6. Conclusion: Emerging Technology
The emerging technologies of Azure Intelligent Document, Azure OpenAI,
LlamaIndex, and LangChain offer powerful capabilities for implementing
Retrieval-Augmented Generation systems in engineering design. While
direct applications of these specific tools in engineering literature are still
emerging, the reviewed studies demonstrate clear potential and need for
such technologies.
Azure Intelligent Document addresses the crucial task of digitizing and
extracting information from engineering documents. Azure OpenAI provides
sophisticated language modeling capabilities that can enhance design
ideation and natural language processing in engineering contexts.
LlamaIndex offers efficient data management and retrieval, critical for
handling the large datasets typical in engineering design. LangChain
facilitates the integration of these capabilities into coherent, interactive
systems tailored to engineering design workflows.
As these technologies continue to evolve and find wider adoption in the
engineering community, we can expect to see more direct applications and
case studies demonstrating their impact on reducing project costs,
improving design outcomes, and fostering innovation across various
engineering domains.

2. Methodology
This section details the methodological approach of Retrieval-Augmented
Generation (RAG) in engineering design. We explore the core components of
the RAG framework, discuss advanced techniques that enhance its
capabilities, and examine the tools and technologies that facilitate its
implementation.
2.1 The RAG Framework
The RAG framework consists of three primary components: Retrieve,
Analyze, and Generate. Each component plays a crucial role in leveraging
existing knowledge and creating innovative design solutions.
2.1.1 Retrieve
The Retrieve stage focuses on gathering relevant data and information from
various sources. This process involves:
- Data Sources: Historical design data, design guidelines, design code
provisions, simulation results, academic literature, patents, and
industry standards are among the diverse sources of information
utilized.
- Advanced Retrieval Techniques: Knowledge graphs and machine
learning algorithms are employed to efficiently identify and retrieve
pertinent information. These techniques can understand complex
 relationships between different pieces of information and extract
relevant data even from unstructured sources.
- Semantic Search: Unlike traditional keyword-based searches,
semantic search understands the intent and contextual meaning of
the query, allowing for more accurate and relevant information
retrieval.
2.1.2 Analyze
In the Analyze phase, the retrieved data is evaluated to extract insights and
identify patterns that inform the design process. This stage involves:
- Data Preprocessing: Cleaning, normalizing, and structuring the
retrieved data to ensure consistency and compatibility for analysis.
- Feature Extraction: Identifying key features and parameters that are
most relevant to the design problem at hand.
- Pattern Recognition: Utilizing machine learning algorithms to detect
patterns, trends, and correlations within the data that may not be
immediately apparent to human designers.
- Performance Metric Assessment: Evaluating design parameters
against predefined performance metrics to gauge the potential
effectiveness of different design approaches.
2.1.3 Generate
The Generate stage focuses on producing innovative design solutions based
on the analysis of the retrieved data. This process includes:
- Generative Design Algorithms: Implementing algorithms such as
Prompt Evolution Design Optimization (PEDO) and AutoTRIZ to create
diverse and feasible design options.
- Constraint Satisfaction: Ensuring that generated designs meet all
specified constraints and requirements of the project.
- Optimization: Refining generated designs to maximize performance
across various metrics such as efficiency, cost-effectiveness, and
sustainability.
- Design Iteration: Continuously improving designs through feedback
loops and iterative refinement based on analysis results.
2.2 Advanced Techniques in RAG
Several advanced techniques have been developed to enhance the
capabilities of the RAG framework:
2.2.1 Physics-Informed Geometry-Aware Neural Operator (PI-GANO)
PI-GANO combines physics-informed learning with a geometry encoder to
handle both varying shapes and PDE parameters. This method:
- Improves the accuracy and efficiency of design simulations by
incorporating physical laws directly into the learning process.
- Enables the model to understand and adapt to different geometric
configurations, making it particularly useful for complex engineering
designs.
- Reduces the reliance on extensive datasets by leveraging physical
principles to guide the learning process.
2.2.2 Physics-Informed Geometric Operators (GOs)
GOs are designed to extract high-level geometric information and physics
from shapes, enhancing the performance of surrogate models. Key features
include:
- Efficient encoding of complex geometric information, allowing for
faster and more accurate analysis of design variations.
- Reduction of overfitting by focusing on physically relevant geometric
features.
- Improved generalization to new design configurations, enabling more
robust design exploration.
2.2.3 Multi-Fidelity Cross-validation (MFCV)
MFCV is a technique that combines multiple fidelity models into a single
surrogate model. This approach:
- Enables more efficient use of computational resources by
strategically balancing high-fidelity and low-fidelity simulations.
- Improves overall accuracy by leveraging the strengths of different
model fidelities.
- Facilitates faster design iterations while maintaining a high level of
confidence in the results.
2.3 Tools and Technologies
The implementation of RAG in engineering design is facilitated by several
cutting-edge tools and technologies:
2.3.1 Azure Intelligent Document
This cloud-based service automates the extraction of relevant data from
various document formats, streamlining the information retrieval process.
Features include:
- Optical Character Recognition (OCR) for converting images and
scanned documents into machine-readable text.
- Natural Language Processing (NLP) capabilities for understanding
document context and extracting key information.
- Integration with other Azure services for seamless data processing
and analysis.
2.3.2 Azure OpenAI
Azure OpenAI provides access to powerful language models for natural
language processing and design ideation. Key capabilities include:
- Text generation and completion for creating design descriptions and
specifications.
- Semantic similarity analysis for comparing different design concepts.
- Integration with other Azure services for end-to-end AI-powered
design workflows.
2.3.3 LlamaIndex
LlamaIndex is a data management tool that enables efficient retrieval of
relevant information from large datasets. It offers:
- Advanced indexing techniques for quick and accurate information
retrieval.
- Customizable query processing to tailor information retrieval to
specific engineering domains.
- Seamless integration with various data sources and formats
commonly used in engineering design.
2.3.4 LangChain
LangChain is a framework for building applications that utilize language
models, facilitating the creation of interactive design tools. It provides:
- A modular architecture for combining different language model
capabilities.
 - Tools for managing context and memory in language model
interactions, crucial for maintaining coherence in complex design
tasks.
- Integration capabilities with external tools and data sources,
enhancing the versatility of RAG applications.
2.3.5 Gradio
Gradio is an open-source library for creating user-friendly interfaces for
machine learning models, enhancing the accessibility of RAG-based design
tools. It offers:
- Rapid prototyping of interactive interfaces for design tools.
- Easy integration with various machine learning models and
frameworks.
- Customizable components for visualizing design outputs and
facilitating user interaction.
By leveraging these advanced techniques and tools, the RAG framework
provides a powerful methodology for enhancing the engineering design
process. The synergy between sophisticated data retrieval, analysis, and
generation capabilities enables engineers to explore design spaces more
efficiently and innovatively than ever before.

3. Results

4. Discussion

5. Conclusion
Retrieval-augmented generation represents a paradigm shift in engineering
design, offering unprecedented opportunities for enhancing efficiency,
accuracy, and innovation. By leveraging advanced AI techniques and
computational tools, RAG enables engineers to tackle complex design
challenges with greater speed and precision than ever before.
As the field continues to evolve, future research should focus on addressing
the challenges identified, particularly in areas of data quality, computational
efficiency, and seamless integration with existing engineering workflows.
Additionally, the development of more intuitive user interfaces and the
expansion of RAG techniques into emerging engineering domains present
exciting opportunities for further innovation.
In conclusion, Retrieval-Augmented Generation stands poised to
revolutionize the field of engineering design, driving innovation and
efficiency across industries. As researchers and practitioners continue to
refine and expand upon these techniques, the future of engineering design
looks brighter and more innovative than ever before.

References
1. Zhong, W., & Meidani, H. (2024). Physics-Informed Geometry-Aware
Neural Operator. Https://arxiv.org/abs/2408.01600
2. Chong, L., Rayan, J., Dow, S., Lykourentzou, I., & Ahmed, F. (2024).
CAD-Prompted Generative Models: A Pathway to Feasible and Novel
Engineering Designs. Https://arxiv.org/abs/2407.08675
3. Khan, S., Masood, Z., Usama, M., Kostas, K., & Kaklis, P. (2024).
Physics-Informed Geometric Operators to Support Surrogate,
Dimension Reduction and Generative Models for Engineering Design.
Https://arxiv.org/abs/2407.07611
4. Renganathan, S. A., & Carlson, K. (2024). Multifidelity Cross-
validation. Https://arxiv.org/abs/2407.01495
5. Tong, D., Hao, Z., & Huang, W. (2024). Inverse Design of Planar
Clamped-Free Elastic Rods from Noisy Data.
Https://arxiv.org/abs/2406.15166
6. Zhou, R., Yuan, C., Permenter, F., Zhang, Y., Arechiga, N., Klenk, M., &
Ahmed, F. (2024). Bridging Design Gaps: A Parametric Data
Completion Approach With Graph Guided Diffusion Models.
Https://arxiv.org/abs/2406.11934
7. Wong, M., Rios, T., Menzel, S., & Ong, Y. S. (2024). Generative AI-based
Prompt Evolution Engineering Design Optimization With Vision-
Language Model. Https://arxiv.org/abs/2406.09143
8. Hong, S., Kwon, Y., Shin, D., Park, J., & Kang, N. (2024). DeepJEB: 3D
Deep Learning-based Synthetic Jet Engine Bracket Dataset.
Https://arxiv.org/abs/2406.09047
9. Park, J., & Kang, N. (2024). Bayesian Mesh Optimization for Graph
Neural Networks to Enhance Engineering Performance Prediction.
Https://arxiv.org/abs/2406.01996
10. Nobari, A. H., Srivastava, A., Gutfreund, D., Xu, K., & Ahmed, F.
(2024). LInK: Learning Joint Representations of Design and
Performance Spaces through Contrastive Learning for Mechanism
Synthesis. Https://arxiv.org/abs/2405.20592
11. Abueidda, D. W., Pantidis, P., & Mobasher, M. E. (2024).
DeepOKAN: Deep Operator Network Based on Kolmogorov Arnold
Networks for Mechanics Problems. Https://arxiv.org/abs/2405.19143
12. Zhang, Y. (2024). Non-Driving-Related Tasks Influencing Drivers'
Takeover Time: A Meta-Analysis. Https://arxiv.org/abs/2405.18667
13. Cheng, Z., Hao, Z., Wang, X., Huang, J., Wu, Y., Liu, X., Zhao, Y.,
Liu, S., & Su, H. (2024). Reference Neural Operators: Learning the
Smooth Dependence of Solutions of PDEs on Geometric
Deformations. Https://arxiv.org/abs/2405.17509
14. Chu, M., & Qian, W. (2024). Physics Constrained Deep Learning
For Turbulence Model Uncertainty Quantification.
Https://arxiv.org/abs/2405.16554
15. Patel, Y., Rayan, S., & Tewari, A. (2024). Conformal Robust
Control of Linear Systems. Https://arxiv.org/abs/2405.16250
16. Zhu, Q., Wang, D., Ma, S., Wang, A. Y., Chen, Z., Khurana, U., &
Ma, X. (2024). Towards Feature Engineering with Human and AI's
Knowledge: Understanding Data Science Practitioners' Perceptions in
Human&AI-Assisted Feature Engineering Design.
https://doi.org/10.1145/3643834.3661517
17. Serani, A., & Diez, M. (2024). A Survey on Design-space
Dimensionality Reduction Methods for Shape Optimization.
Https://arxiv.org/abs/2405.13944
18. Mayeur, J., Cheng, J., Lee, Y., Knipe, K., & Kadau, K. (2024). Self-
supervised feature distillation and design of experiments for efficient
training of micromechanical deep learning surrogates.
Https://arxiv.org/abs/2405.10135
19. Chu, M., & Qian, W. (2024). A Deep Learning Approach For
Epistemic Uncertainty Quantification Of Turbulent Flow Simulations.
Https://arxiv.org/abs/2405.08148
20. Okpala, E., Vishwamitra, N., Guo, K., Liao, S., Cheng, L., Hu, H.,
Wu, Y., Yuan, X., Wade, J., & Khorsandroo, S. (2024). AI-Cybersecurity
Education Through Designing AI-based Cyberharassment Detection
Lab. Https://arxiv.org/abs/2405.08125
21. Hamade, K., Sen, S., Kleinberg, J., & Anderson, A. (2024).
Designing Skill-Compatible AI: Methodologies and Frameworks in
Chess. Https://arxiv.org/abs/2405.05066
22. Arisaka, S., & Li, Q. (2024). Accelerating Legacy Numerical
Solvers by Non-intrusive Gradient-based Meta-solving.
Https://arxiv.org/abs/2405.02952
23. Ege, D. N., Øvrebø, H. H., Stubberud, V., Berg, M. F., Elverum, C.,
Steinert, M., & Vestad, H. (2024). ChatGPT as an inventor: Eliciting the
strengths and weaknesses of current large language models against
humans in engineering design. Https://arxiv.org/abs/2404.18479
24. Lou, T., Guan, G., Yue, Z., Wang, Y., Qi, R., & Tong, S. (2024). A
competitive game optimization algorithm for Unmanned Aerial
Vehicle path planning. Https://arxiv.org/abs/2404.09567
25. Doris, A. C., Grandi, D., Tomich, R., Alam, M. F., Cheong, H., &
Ahmed, F. (2024). DesignQA: A Multimodal Benchmark for Evaluating
Large Language Models' Understanding of Engineering
Documentation. Https://arxiv.org/abs/2404.07917
26. Picard, C., & Ahmed, F. (2024). Fast and Accurate Bayesian
Optimization with Pre-trained Transformers for Constrained
Engineering Problems. Https://arxiv.org/abs/2404.04495
27. Kevian, D., Syed, U., Guo, X., Havens, A., Dullerud, G., Seiler, P.,
Qin, L., & Hu, B. (2024). Capabilities of Large Language Models in
Control Engineering: A Benchmark Study on GPT-4, Claude 3 Opus,
and Gemini 1.0 Ultra. Https://arxiv.org/abs/2404.03647
28. He, J., Koric, S., Abueidda, D., Najafi, A., & Jasiuk, I. (2024).
Geom-DeepONet: A Point-cloud-based Deep Operator Network for
Field Predictions on 3D Parameterized Geometries.
https://doi.org/10.1016/j.cma.2024.117130
29. Wang, J., Petra, C. G., & Peterson, J. L. (2024). Constrained
Bayesian optimization with merit functions.
Https://arxiv.org/abs/2403.13140
30. Jiang, S., & Luo, J. (2024). AutoTRIZ: Artificial Ideation with TRIZ
and Large Language Models. Https://arxiv.org/abs/2403.13002
31. Liu, P., Hong, Y., Escobar, L. A., & Meeker, W. Q. (2024). On
Equivalence of Likelihood-Based Confidence Bands for Fatigue-Life
and Fatigue-Strength Distributions. Https://arxiv.org/abs/2403.12757
32. Zhao, Z., Liu, C., Li, Y., Chen, Z., & Liu, X. (2024). Diffeomorphism
Neural Operator for various domains and parameters of partial
differential equations. Https://arxiv.org/abs/2402.12475
33. Li, K., & Li, F. (2024). Multi-Fidelity Methods for Optimization: A
Survey. Https://arxiv.org/abs/2402.09638
34. Usama, M., Masood, Z., Khan, S., Kostas, K., & Kaklis, P. (2024).
Generative VS non-Generative Models in Engineering Shape
Optimization. Https://arxiv.org/abs/2402.08540
35. Cabral, R., Kalinowski, M., Baldassarre, M. T., Villamizar, H.,
Escovedo, T., & Lopes, H. (2024). Investigating the Impact of SOLID
Design Principles on Machine Learning Code Understanding.
Https://arxiv.org/abs/2402.05337
36. Nobari, A. H., Giannone, G., Regenwetter, L., & Ahmed, F. (2024).
NITO: Neural Implicit Fields for Resolution-free Topology
Optimization. Https://arxiv.org/abs/2402.05073
37. Kim, H. (2024). Enhancing Gaussian Process Surrogates for
Optimization and Posterior Approximation via Random Exploration.
Https://arxiv.org/abs/2401.17037
38. Lou, Y., Sun, S., & Nocedal, J. (2024). Noise-Tolerant
Optimization Methods for the Solution of a Robust Design Problem.
Https://arxiv.org/abs/2401.15007
39. Porter, J. H., & Brake, M. R. (2024). Tracking Superharmonic
Resonances for Nonlinear Vibration of Conservative and Hysteretic
Single Degree of Freedom Systems.
https://doi.org/10.1016/j.ymssp.2024.111410
40. Naeini, S. S., & Snaiki, R. (2024). A Physics-informed machine
learning model for time-dependent wave runup prediction.
Https://arxiv.org/abs/2401.08684
41. Peng, Y., Niloy, I., Kam, M., Celli, P., & Plucinsky, P. (2024).
Programming bistability in geometrically perturbed mechanical
metamaterials. Https://arxiv.org/abs/2401.07881
42. Picard, C., & Ahmed, F. (2024). Fast and Accurate Zero-Training
Classification for Tabular Engineering Data.
Https://arxiv.org/abs/2401.06948
43. Study Group, T. C. (2023). CEPC Technical Design Report --
Accelerator (v2). Https://arxiv.org/abs/2312.14363
44. Yang, L., Zhang, J., Li, H., Ren, L., Yang, C., Wang, J., & Shi, D.
(2023). A Comprehensive End-to-End Computer Vision Framework for
Restoration and Recognition of Low-Quality Engineering Drawings.
Https://arxiv.org/abs/2312.13620
45. Emuna, H., Borenstein, N., Qian, X., Kang, H., Chan, J., Kittur, A.,
& Shahaf, D. (2023). Imitation of Life: A Search Engine for Biologically
Inspired Design. Https://arxiv.org/abs/2312.12681
46. Namura, N. (2023). Single and Multi-Objective Optimization
Benchmark Problems Focusing on Human-Powered Aircraft Design.
Https://arxiv.org/abs/2312.08953
47. Zhang, J., Chen, X., Wang, Y., Wang, C., Liu, Y., Li, X., Yang, M., &
Tao, D. (2023). Exploring Plain ViT Reconstruction for Multi-class
Unsupervised Anomaly Detection. Https://arxiv.org/abs/2312.07495
48. Siddharth, L., & Luo, J. (2023). Linguistic and Structural Basis of
Engineering Design Knowledge. Https://arxiv.org/abs/2312.06355
49. Wang, Z., Xiong, J., Liu, X., Liu, L., & Lai, X. (2023). An Updated
Lagrangian Particle Hydrodynamics (ULPH)-NOSBPD Coupling
Approach for Modeling Fluid-Structure Interaction Problems.
Https://arxiv.org/abs/2312.05016
50. Hawthorne, F., Cleuren, B., & Fiore, C. E. (2023).
Thermodynamics of a minimal collective heat engine: Comparison
between engine designs. Https://arxiv.org/abs/2311.18629
51. Jessica, L. S., Arafat, N. A., Lim, W. X., Chan, W. L., & Kong, A. W.
(2023). Finite Volume Features, Global Geometry Representations,
and Residual Training for Deep Learning-based CFD Simulation.
Https://arxiv.org/abs/2311.14464
52. Do, B., & Zhang, R. (2023). Multi-fidelity Bayesian Optimization
in Engineering Design. Https://arxiv.org/abs/2311.13050
53. Picard, C., Edwards, K. M., Doris, A. C., Man, B., Giannone, G.,
Alam, M. F., & Ahmed, F. (2023). From Concept to Manufacturing:
Evaluating Vision-Language Models for Engineering Design.
Https://arxiv.org/abs/2311.12668
54. Wang, Z., Mantey, S. K., & Zhang, X. (2023). A numerical tool for
efficient analysis and optimization of offshore wind turbine jacket
substructure considering realistic boundary and loading conditions.
Https://arxiv.org/abs/2310.19388
55. Lie, H., Studer, K., Zhao, Z., Thomson, B., Turakhia, D. G., & Liu,
J. (2023). Training for Open-Ended Drilling through a Virtual Reality
Simulation. Https://arxiv.org/abs/2310.17417
56. Azizzadenesheli, K., Kovachki, N., Li, Z., Kossaifi, J., &
Anandkumar, A. (2023). Neural Operators for Accelerating Scientific
Simulations and Design. Https://arxiv.org/abs/2309.15325
57. Kang, R., Mu, T., Liatsis, P., & Kyritsis, D. C. (2023). Physics-
Driven ML-Based Modelling for Correcting Inverse Estimation.
Https://arxiv.org/abs/2309.13985
58. Han, T., Li, J., Guo, Z., & Jin, Y. (2023). A Region-Shrinking-Based
Acceleration for Classification-Based Derivative-Free Optimization.
Https://arxiv.org/abs/2309.11036
59. Zhao, Y., Liu, Y., & Xu, Z. (2023). Predicting Fatigue Crack Growth
via Path Slicing and Re-Weighting.
https://doi.org/10.1016/j.taml.2023.100477
60. Sáez, A., & Lecampion, B. (2023). Fluid-driven slow slip and
earthquake nucleation on a slip-weakening circular fault.
Https://arxiv.org/abs/2309.04567
61. Li, C., Chen, H., Yan, M., Shen, W., Xu, H., Wu, Z., Zhang, Z.,
Zhou, W., Chen, Y., Cheng, C., Shi, H., Zhang, J., Huang, F., & Zhou, J.
(2023). ModelScope-Agent: Building Your Customizable Agent System
with Open-source Large Language Models.
Https://arxiv.org/abs/2309.00986
62. Goyal, P., Duff, I. P., & Benner, P. (2023). Guaranteed Stable
Quadratic Models and their applications in SINDy and Operator
Inference. Https://arxiv.org/abs/2308.13819
63. Jwa, M., Kim, J., Shin, S., Jin, A., Shin, D., & Kang, N. (2023).
Performance Comparison of Design Optimization and Deep Learning-
based Inverse Design. Https://arxiv.org/abs/2308.13000
64. Girija, A. P. (2023). Planetary Entry Probe Dataset: Analysis and
Rules of Thumb for Future Missions. Https://arxiv.org/abs/2308.07005
65. Constantinides, M., Bogucka, E., Quercia, D., Kallio, S., &
Tahaei, M. (2023). RAI Guidelines: Method for Generating Responsible
AI Guidelines Grounded in Regulations and Usable by (Non-)Technical
Roles. Https://arxiv.org/abs/2307.15158
66. Fan, J., Vuaille, L., Wang, H., & Bäck, T. (2023). Adversarial Latent
Autoencoder with Self-Attention for Structural Image Synthesis.
https://doi.org/10.1109/CAI59869.2024.00030
67. Siddharth, L., & Luo, J. (2023). Retrieval Augmented Generation
using Engineering Design Knowledge.
Https://arxiv.org/abs/2307.06985
68. Rios, T., Menzel, S., & Sendhoff, B. (2023). Large Language and
Text-to-3D Models for Engineering Design Optimization.
Https://arxiv.org/abs/2307.01230
69. Giannone, G., Regenwetter, L., Srivastava, A., Gutfreund, D., &
Ahmed, F. (2023). Learning from Invalid Data: On Constraint
Satisfaction in Generative Models. Https://arxiv.org/abs/2306.15166
70. Szaksz, B., Stepan, G., & Habib, G. (2023). Dynamical integrity
estimation in time delayed systems: A rapid iterative algorithm.
Https://arxiv.org/abs/2306.14821
71. Saxena, N., Jain, M., Liu, C., & Bengio, Y. (2023). Multi-Fidelity
Active Learning with GFlowNets. Https://arxiv.org/abs/2306.11715
72. Arechiga, N., Permenter, F., Song, B., & Yuan, C. (2023). Drag-
guided diffusion models for vehicle image generation.
Https://arxiv.org/abs/2306.09935
73. Göpfert, J., Weinand, J. M., Kuckertz, P., & Stolten, D. (2023).
Opportunities for Large Language Models and Discourse in
Engineering Design. Https://arxiv.org/abs/2306.09169
74. Song, B., Yuan, C., Permenter, F., Arechiga, N., & Ahmed, F.
(2023). Surrogate Modeling of Car Drag Coefficient with Depth and
Normal Renderings. Https://arxiv.org/abs/2306.06110
75. Cobb, A. D., Roy, A., Elenius, D., Heim, F. M., Swenson, B.,
Whittington, S., Walker, J. D., Bapty, T., Hite, J., Ramani, K., McComb,
C., & Jha, S. (2023). AircraftVerse: A Large-Scale Multimodal Dataset
of Aerial Vehicle Designs. Https://arxiv.org/abs/2306.05562
76. Ansari, N., Seidel, H., & Babaei, V. (2023). Large-Batch, Iteration-
Efficient Neural Bayesian Design Optimization.
Https://arxiv.org/abs/2306.01095
77. Picard, C., Schiffmann, J., & Ahmed, F. (2023). DATED:
Guidelines for Creating Synthetic Datasets for Engineering Design
Applications. Https://arxiv.org/abs/2305.09018
78. Wang, R., Shapiro, V., & Behandish, M. (2023). Model
Consistency for Mechanical Design: Bridging Lumped and Distributed
Parameter Models with A Priori Guarantees.
https://doi.org/10.1115/1.4064810
79. Seurin, P., & Shirvan, K. (2023). Assessment of Reinforcement
Learning Algorithms for Nuclear Power Plant Fuel Optimization.
Https://arxiv.org/abs/2305.05812
80. Nemani, V., Biggio, L., Huan, X., Hu, Z., Fink, O., Tran, A., Wang,
Y., Zhang, X., & Hu, C. (2023). Uncertainty Quantification in Machine
Learning for Engineering Design and Health Prognostics: A Tutorial.
https://doi.org/10.1016/j.ymssp.2023.110796
81. Bollapragada, R., Karamanli, C., Keith, B., Lazarov, B., Petrides,
S., & Wang, J. (2023). An Adaptive Sampling Augmented Lagrangian
Method for Stochastic Optimization with Deterministic Constraints.
Https://arxiv.org/abs/2305.01018
82. Alpat, A. B., Coban, A., Kaya, H., & Bartolini, G. (2023).
MRADSIM-Converter: A new software for STEP to GDML conversion.
https://doi.org/10.1016/j.cpc.2023.108688
83. Sharma, G., Mun, S., Lee, N., Peterson, L., Tellkamp, D., &
Nellippallil, A. B. (2023). Machine Learning-Based Multi-Objective
Design Exploration Of Flexible Disc Elements.
Https://arxiv.org/abs/2304.07245
84. Chen, L., Yao, X., Liu, K., Tan, C., & Moon, S. K. (2023).
Multisensor fusion-based digital twin in additive manufacturing for in-
situ quality monitoring and defect correction.
Https://arxiv.org/abs/2304.05685
85. Mallon, C. N., Thornton, A. W., Hill, M. R., & Badia, S. (2023).
Neural Level Set Topology Optimization Using Unfitted Finite
Elements. Https://arxiv.org/abs/2303.13672
86. Ahmadianshalchi, A., Belakaria, S., & Doppa, J. R. (2023).
Preference-Aware Constrained Multi-Objective Bayesian
Optimization. Https://arxiv.org/abs/2303.13034
87. Kochiyama, S., & Kesari, H. (2023). Functional significance of
lamellar architecture in marine sponge fibers: Conditions for when
splitting a cylindrical tube into an assembly of tubes will decrease its
bending stiffness. Https://arxiv.org/abs/2303.12189-
88. Giannone, G., & Ahmed, F. (2023). Diffusing the Optimal
Topology: A Generative Optimization Approach.
Https://arxiv.org/abs/2303.09760
89. Kumar, S., & Atre, S. V. (2023). Design and optimization of brake
disc using Multi-Objective genetic algorithm.
https://doi.org/10.46720/F2020-ADM-071
90. Han, R., Kramer, B., Lee, D., Narayan, A., & Xu, Y. (2023). An
approximate control variates approach to multifidelity distribution
estimation. Https://arxiv.org/abs/2303.06422
91. Gmeiner, F., Yang, H., Yao, L., Holstein, K., & Martelaro, N.
(2023). Exploring Challenges and Opportunities to Support Designers
in Learning to Co-create with AI-based Manufacturing Design Tools.
https://doi.org/10.1145/3544548.3580999
92. Vardhan, H., Volgyesi, P., & Sztipanovits, J. (2023). Fusion of ML
with numerical simulation for optimized propeller design.
Https://arxiv.org/abs/2302.14740
93. Toklu, N. E., Atkinson, T., Micka, V., Liskowski, P., & Srivastava,
R. K. (2023). EvoTorch: Scalable Evolutionary Computation in Python.
Https://arxiv.org/abs/2302.12600
94. Song, B., Zhou, R., & Ahmed, F. (2023). Multi-modal Machine
Learning in Engineering Design: A Review and Future Directions.
Https://arxiv.org/abs/2302.10909
95. Regenwetter, L., Srivastava, A., Gutfreund, D., & Ahmed, F.
(2023). Beyond Statistical Similarity: Rethinking Metrics for Deep
Generative Models in Engineering Design.
Https://arxiv.org/abs/2302.02913
96. Yang, M., Jiang, P., Zang, T., & Liu, Y. (2023). Data-driven
intelligent computational design for products: Method, techniques,
and applications. Https://arxiv.org/abs/2301.12382
97. Goyal, P., Duff, I. P., & Benner, P. (2023). Inference of Continuous
Linear Systems from Data with Guaranteed Stability.
Https://arxiv.org/abs/2301.10060
98. Segarra, S., Paoletti, J., Zajac, S., Beier, M. E., Sabharwal, A.,
Wettergreen, M., & Salas, E. (2023). Conversational Turn-taking as a
Stochastic Process on Networks. Https://arxiv.org/abs/2301.04030
99. Okada, A., Yoshida, H., Kidono, K., Matsumori, T., Takeno, T., &
Kadowaki, T. (2023). Design Optimization of Noise Filter using
Quantum Annealer. https://doi.org/10.1109/ACCESS.2023.3271969
100. Zhang, H., Zhou, Y., & Li, F. (2023). An Automatic Method for
Generating Symbolic Expressions of Zernike Circular Polynomials.
https://doi.org/10.1109/ACCESS.2023.3283028
101. Huang, J. (2023). Digital Engineering Transformation with
Trustworthy AI towards Industry 4.0: Emerging Paradigm Shifts.
Https://arxiv.org/abs/2301.00951
102. Arora, R., & Shrivastava, A. (2022). Spatio-Temporal Super-
Resolution of Dynamical Systems using Physics-Informed Deep-
Learning. Https://arxiv.org/abs/2212.04457
103. Bond, R. B., Ren, P., Hajjar, J. F., & Sun, H. (2022). An
Unsupervised Machine Learning Approach for Ground-Motion Spectra
Clustering and Selection. https://doi.org/10.1002/eqe.4062
104. Fan, J., Vuaille, L., Wang, H., & Bäck, T. (2023). Adversarial Latent
Autoencoder with Self-Attention for Structural Image Synthesis.
https://doi.org/10.1109/CAI59869.2024.00030
105. Zhang, W., Joseph, J., Yin, Y., Xie, L., Furuhata, T., Yamakawa, S.,
Shimada, K., & Kara, L. B. (2022). Component Segmentation of
Engineering Drawings Using Graph Convolutional Networks.
https://doi.org/10.1016/j.compind.2023.103885
106. Raina, A., Cagan, J., & McComb, C. (2022). Learning to design
without prior data: Discovering generalizable design strategies using
deep learning and tree search. https://doi.org/10.1115/1.4056221
107. Vardhan, H., Timalsina, U., Volgyesi, P., & Sztipanovits, J. (2022).
Data efficient surrogate modeling for engineering design: Ensemble-
free batch mode deep active learning for regression.
Https://arxiv.org/abs/2211.10360
108. Hao, Z., Liu, S., Zhang, Y., Ying, C., Feng, Y., Su, H., & Zhu, J.
(2022). Physics-Informed Machine Learning: A Survey on Problems,
Methods and Applications. Https://arxiv.org/abs/2211.08064
109. Ardalani, N., Pal, S., & Gupta, P. (2022). DeepFlow: A Cross-
Stack Pathfinding Framework for Distributed AI Systems.
Https://arxiv.org/abs/2211.03309
110. Kumar, D., Ahmed, F., Usman, S., Alajo, A., & Alam, S. (2022).
Recent Advances in Uncertainty Quantification Methods for
Engineering Problems. Https://arxiv.org/abs/2211.03012
111. Siddharth, L., Li, G., & Luo, J. (2022). Enhancing Patent Retrieval
using Text and Knowledge Graph Embeddings: A Technical Note.
Https://arxiv.org/abs/2211.01976
112. Casey, S. (2022). Quantifying Complexity: An Object-Relations
Approach to Complex Systems. Https://arxiv.org/abs/2210.12347
113. Wei, H., Lyu, D., Hu, W., & Wu, C. T. (2022). RVE Analysis in LS-
DYNA for High-fidelity Multiscale Material Modeling.
Https://arxiv.org/abs/2210.11761
114. Filanova, Y., Duff, I. P., Goyal, P., & Benner, P. (2022). An Operator
Inference Oriented Approach for Mechanical Systems.
https://doi.org/10.1016/j.ymssp.2023.110620
115. Sadeghi, J., Mueller, R., & Redford, J. (2022). An Active Learning
Reliability Method for Systems with Partially Defined Performance
Functions. Https://arxiv.org/abs/2210.02168
116. Mattioli, J., Delaborde, A., Khalfaoui, S., Lecue, F., Sohier, H., &
Jurie, F. (2022). Empowering the trustworthiness of ML-based critical
systems through engineering activities.
Https://arxiv.org/abs/2209.15438
117. Song, Z., Yao, H., Tian, D., & Zhan, G. (2022). Cyclegan Network
for Sheet Metal Welding Drawing Translation.
Https://arxiv.org/abs/2209.14106
118. Song, Z., & Yao, H. (2022). Segmentation method of U-net sheet
metal engineering drawing based on CBAM attention mechanism.
Https://arxiv.org/abs/2209.14102
119. Ghosh, A., Deb, K., Goodman, E., & Averill, R. (2022). An
Interactive Knowledge-based Multi-objective Evolutionary Algorithm
Framework for Practical Optimization Problems.
Https://arxiv.org/abs/2209.08604
120. Vani, N., Escudier, S., & Sauret, A. (2022). Influence of the solid
fraction on the clogging by bridging of suspensions in constricted
channels. Https://arxiv.org/abs/2208.11643
121. Mazé, F., & Ahmed, F. (2022). Diffusion Models Beat GANs on
Topology Optimization. Https://arxiv.org/abs/2208.09591
122. Arora, R. (2022). PhySRNet: Physics informed super-resolution
network for application in computational solid mechanics.
Https://arxiv.org/abs/2206.15457
123. Vardhan, H., & Sztipanovits, J. (2022). Deep Active Learning for
Regression Using $\epsilon$-weighted Hybrid Query Strategy.
Https://arxiv.org/abs/2206.13298
124. Chen, Z., Li, Q., & Zhang, Z. (2022). Self-Healing Robust Neural
Networks via Closed-Loop Control. Https://arxiv.org/abs/2206.12963
125. Thomas, S. S., Palandri, J., Chakravarty, P., & Blaschko, M. B.
(2022). Designing MacPherson Suspension Architectures using
Bayesian Optimization. Https://arxiv.org/abs/2206.09022
126. Regenwetter, L., & Ahmed, F. (2022). Towards Goal, Feasibility,
and Diversity-Oriented Deep Generative Models in Design.
Https://arxiv.org/abs/2206.07170
127. Khoshoo, B., Blank, J., Pham, T. Q., Deb, K., & Foster, S. N. (2022).
Optimal Design of Electric Machine with Efficient Handling of
Constraints and Surrogate Assistance.
Https://arxiv.org/abs/2206.01695
128. Sarkar, S., Pandey, P., & Kar, S. (2022). Automatic Detection and
Classification of Symbols in Engineering Drawings.
Https://arxiv.org/abs/2204.13277
129. Andersen, B., Delipei, G., Kropaczek, D., & Hou, J. (2022). MOF:
A Modular Framework for Rapid Application of Optimization
Methodologies to General Engineering Design Problems.
Https://arxiv.org/abs/2204.00141
130. Li, Y., & Dankowicz, H. (2022). Model-free Continuation of
Periodic Orbits in Certain Nonlinear Systems Using Continuous-Time
Adaptive Control. https://doi.org/10.1007/s11071-022-08059-1
131. Regenwetter, L., Weaver, C., & Ahmed, F. (2022). FRAMED: An
AutoML Approach for Structural Performance Prediction of Bicycle
Frames. Https://arxiv.org/abs/2201.10459
132. Karcher, C., & Haimes, R. (2022). A Method of Sequential Log-
Convex Programming for Engineering Design.
Https://arxiv.org/abs/2201.08436
133. Astudillo, R., & Frazier, P. I. (2021). Bayesian Optimization of
Function Networks. Https://arxiv.org/abs/2112.15311
 134. Knaster, J., Ibarra, A., Abal, J., Sena, A. A., Arbeiter, F., Arranz, F.,
Arroyo, J. M., Bargallo, E., Beauvais, P. Y., Bernardi, D., Casal, N.,
Carmona, J. M., Chauvin, N., Comunian, M., Delferriere, O., Delgado,
A., Arocas, P. D., Fischer, U., Frisoni, M., . . . Yokomine, T. (2021). The
accomplishment of the Engineering Design Activities of IFMIF/EVEDA:
The European Japanese project towards a Li(d,xn) fusion relevant
neutron source. https://doi.org/10.1088/0029-5515/55/8/086003
135. Lambert, J., Liu, Z., Sener, O., Hays, J., & Koltun, V. (2021). MSeg:
A Composite Dataset for Multi-domain Semantic Segmentation.
Https://arxiv.org/abs/2112.13762
136. Harish, A. B., & Prasad, A. R. (2021). Photo2CAD: Automated 3D
solid reconstruction from 2D drawings using OpenCV.
Https://arxiv.org/abs/2101.04248
137. Chandrasekhar, A., Sridhara, S., & Suresh, K. (2021). Integrating
Material Selection with Design Optimization via Neural Networks.
Https://arxiv.org/abs/2112.12566
138. https://azure.microsoft.com/en-in/products/ai-services/ai-
document-intelligence
139. https://learn.microsoft.com/en-us/azure/ai-
services/openai/overview
140. https://www.llamaindex.ai/
141. https://www.langchain.com/

Appendix A: Physics-Informed Geometry-Aware Neural Operator (PI-GANO)

This appendix provides a detailed overview of the Physics-Informed
Geometry-Aware Neural Operator (PI-GANO), including its architecture, key
components, and training methodology.
A.1 Overview
PI-GANO is designed to solve parametric Partial Differential Equations
(PDEs) with varying domain geometries and PDE parameters. It combines
the strengths of Physics-Informed Deep Compositional Operator Network
(PI-DCON) and Physics-Informed PointNet (PI-PointNet) to create a neural
operator capable of generalizing across both PDE parameters and domain
geometries without requiring training data.
A.2 Problem Setting
The goal is to develop an efficient machine learning-based solver for
parametric PDEs formulated as:
N_x[u(x), k(x)] = 0, x ∈ Ω,
B_x[u(x)] = g(x), x ∈ ∂Ω,
where:
- Ω is a physical domain in R^d
- x is a d-dimensional vector of spatial coordinates
- N_x is a general differential operator
- B_x is a boundary condition operator acting on the domain boundary
∂Ω
- k(x) refers to the parameters of the PDE
- g(x) denotes the boundary conditions
- u(x) is the solution of the PDE
The aim is to approximate the operator M defined by:
M: {k_i(x), g_i(x), Ω_i} → u_i(x), x ∈ Ω_i, Ω_i ∈ Ω
A.3 Model Architecture
PI-GANO's architecture consists of several key components:
A.3.1 Geometry Encoder
The geometry encoder captures the domain geometry features:
1. Represent domain geometry Ω_i using a set of collocation points
{x_j^i}_{j=1}^{s_i}.
2. Compute high-dimensional features using an MLP mapping U_G(·,
θ_G).
3. Apply average pooling to extract global geometry features:
G_i = AvgPool({h_j}_{j=1}^{s_i}),
where {h_j}_{j=1}^{s_i} = {U_G(x_j, θ_G)}_{j=1}^{s_i}
A.3.2 Parameter Encoder
The parameter encoder processes the PDE parameters:
1. Represent PDE parameters in discrete form by evaluating functions at
sampled coordinates.
 2. Apply a multi-layer perceptron U_b(·, θ_b) with a Max-pooling layer:
b = Maxpool({U_b(g(x'_j), θ_b)}_{j=1}^m)
A.3.3 Local Coordinate Feature Computation
Compute local coordinate features for each collocation point:
h_i = σ(W_t x_i + B_t)
A.3.4 Feature Combination
Concatenate the geometry global feature G_i with each local coordinate
feature h_i:
H_i = [h_i ∥ G_i]
A.3.5 Operator Layers
Apply a series of operator layers to produce the PDE solution:
u_i(x_j^i) ≈ M_θ(x_j^i, g_i(x), Ω_i) = sum{b ⊙ ...(b ⊙ (W_O^2 σ(b ⊙
(W_O^1 [σ(W_t x_i + B_t)∥G_i] + B_O^1)) + B_O^2))}
A.4 Training Methodology
PI-GANO is trained using a physics-informed approach, which incorporates
the governing PDEs directly into the loss function:
L(θ) = L_PDE(θ) + αL_BC(θ)
where:
- L_PDE(θ) is the PDE residual loss
- L_BC(θ) is the boundary condition loss
- α is a weighting hyperparameter
The specific formulations of L_PDE(θ) and L_BC(θ) depend on the problem
being solved. For example, in the Darcy flow problem:
L_PDE(θ) = (1/N) ∑_{i=1}^N (N_x[U_θ(x, k_i(x'), g_i(x'')), k_i(x')])^2, x ∈ Ω
L_BC(θ) = (1/N) ∑_{i=1}^N (B_x[U_θ(x, k_i(x'), g_i(x''))] - g_i(x''))^2, x ∈ ∂Ω
A.5 Key Advantages
1. Generalization: PI-GANO can handle both varying PDE parameters
and domain geometries.
2. Data-free: No need for training data obtained from high-fidelity
simulations.
 3. Efficiency: Faster training compared to data-driven methods,
especially for fine meshes.
4. Accuracy: Improved performance compared to existing physics-
informed neural operators.
A.6 Limitations and Future Work
1. Current focus on steady-state solutions; need to adapt for time-
dependent PDEs.
2. Potential instability in physics-informed training, requiring more
training epochs.
3. Opportunity to incorporate more sophisticated architectures, such as
attention mechanisms.
By addressing these limitations and building upon the strengths of PI-GANO,
future research can further enhance the capabilities of neural operators in
solving complex engineering and scientific problems involving PDEs on
varying geometries.

Appendix B: Physics-Informed Geometric Operators (GOs)

This appendix provides a detailed overview of Physics-Informed Geometric
Operators (GOs), including their mathematical formulation, computation
methods, and their relation to physics in engineering design problems.
B.1 Overview
Physics-Informed Geometric Operators (GOs) are proposed to enrich
geometric data provided to surrogate models, dimension reduction models,
and generative models in engineering design. GOs leverage the shape's
differential and integral properties to capture varying characteristics of the
shape related to its volume distribution, complexity of the bounding surface,
and overall surface smoothness.
B.2 Components of Geometric Operators
The GOs are defined as:
GO(G) = (P(G), M(G), K(G), F(G))
Where:
- G is a geometric object representing a baseline design
- P(G) is a vector function providing a suitable representation of G
- M(G) represents geometric moments
 - K(G) represents curvature integrals
- F(G) represents Fourier descriptors
B.3 Geometric Moments
B.3.1 Definition
An s-order geometric moment Ms = Mp,q,r of a geometric object G is defined
as:
Mp,q,r(G) = ∫∫∫G x^p y^q z^r dx dy dz, with p, q, r ∈ {0, 1, 2, ...}
B.3.2 Computation
Geometric moments are typically computed using Gauss's divergence
theorem, which converts volume integrals to surface integrals.
B.3.3 Relation to Physics
Geometric moments are strongly connected to physical quantities that
determine design performance characteristics. For example, in ship hull
design, the moments of the Sectional Area Curve (SAC) are directly related
to the ship's hydrodynamic properties.
B.4 Curvature
B.4.1 Gaussian Curvature
For a surface P(u,v), the Gaussian curvature K(u,v) is defined as:
K(u,v) = κ1κ2 = (LN - M^2) / (EG - F^2)
Where L, M, N, E, F, G are coefficients of the first and second fundamental
forms of the surface.
B.4.2 Total Curvature
Total curvature is the integral of Gaussian curvature over the entire surface.
It's connected to the Euler characteristic through the Gauss-Bonnet
Theorem.
B.4.3 Relation to Physics
Curvature is linked to fluid flow behavior around objects. For example, in
aerofoil design, curvature distribution affects aerodynamic performance
and heat transfer characteristics.
B.5 Three-dimensional Fourier Descriptor
B.5.1 Definition
Fourier Descriptors (FDs) are particularly sensitive to the geometry of the
shape's boundary. They provide an enhanced description of the object's
boundary surface.
B.5.2 Computation
The computation of 3D FDs is based on the work of Park and Lee, which
involves:
1. Voxelization of the 3D object
2. Computation of the 3D Fourier transform
3. Extraction of Fourier coefficients
B.5.3 Relation to Physics
FDs are particularly useful in capturing the boundary characteristics of
functional surfaces like wings, blades, and ship hulls, which are critical for
their performance.
B.6 Application in Machine Learning Models
GOs can be used to augment training data for various machine learning
models:
1. Surrogate Models: GOs improve the accuracy of performance
prediction models.
2. Dimension Reduction Models: GOs enable the extraction of more
efficient and physically meaningful latent spaces.
3. Generative Models: GOs help in generating more diverse and valid
designs and can be used as a substitute for computationally
expensive physics-based metrics.
B.7 Benefits of GOs
1. Capture both global and local shape features
2. Embed physical information in geometric representations
3. Improve model accuracy and generalization capability
4. Reduce the dimensionality of design spaces while maintaining
diversity and validity
5. Provide a computationally efficient substitute for physics-based
metrics in some applications
By incorporating these physics-informed geometric operators, machine
learning models in engineering design can achieve better performance,
improved generalization, and more physically meaningful results.

Appendix C: CAD-Prompted Generative Models

This appendix provides a detailed overview of CAD-Prompted Generative
Models, including their methodology, implementation, and implications for
engineering design.
C.1 Overview
CAD-Prompted Generative Models are an innovative approach to improving
the feasibility of designs generated by text-to-image (T2I) models in
engineering design contexts. This method addresses the limitation of T2I
models in producing feasible design concepts by incorporating Computer-
Aided Design (CAD) images as prompts alongside text inputs.
C.2 Methodology
C.2.1 Workflow
1. Text Prompt Input: The designer provides a text description of the
desired design.
2. CAD Image Selection: Using OpenAI's CLIP model, the system
identifies a CAD image from a dataset that best matches the text
prompt semantically.
3. Image Generation: Both the text prompt and the selected CAD image
are used as inputs for the T2I model (e.g., Stable Diffusion 2.1).
4. Output: The T2I model generates new design images, blending
characteristics from both the text and CAD image prompts.
C.2.2 CAD Image Dataset
The method relies on a dataset of CAD images relevant to the product being
designed. For example, in a bike design task, the dataset would consist of
diverse bike designs in CAD image format.
C.2.3 Prompt Weighting
The influence of the CAD image on the generated design can be adjusted
through a weighting system. Weights typically range from 0.35 to 1, where
higher weights give more prominence to the CAD image characteristics in
the final output.
C.3 Implementation
C.3.1 Tools and Models
- T2I Model: Stable Diffusion 2.1 (via Leonardo.AI platform)
- Image Selection: OpenAI's CLIP model
- CAD Image Dataset: e.g., BIKED dataset for bike designs
C.3.2 Generation Settings
Multiple settings can be used to explore different balances between text and
CAD image prompts:
1. Stable Diffusion 2.1 (SD)
2. SD + Prompt Magic (SD+PM)
3. SD+PM + CAD Image Prompting (CIP) with varying weights (e.g., 0.35,
0.51, 0.67, 0.83, 1)
C.4 Evaluation Metrics
1. Perceived Feasibility: How manufacturable and realistic the generated
design appears.
2. Novelty: The uniqueness and creativity of the generated design.
3. CLIP Similarity Score: Measure of similarity between designs
generated under different settings.
C.5 Key Findings
1. Improved Feasibility: CAD image prompting significantly enhances the
perceived feasibility of generated designs.
2. Feasibility-Novelty Tradeoff: Higher CAD image prompt weights tend
to increase feasibility but decrease novelty.
3. Optimal Weighting: A CAD image prompt weight around 0.35 can
improve feasibility without significantly compromising novelty.
4. Weight Threshold: Weights above 0.83 may lead to decreased
feasibility, possibly due to the model's difficulty in balancing text and
image inputs.
C.6 Implications for Engineering Design
1. Expanded Applicability: Enables T2I models to generate more realistic
and potentially manufacturable designs.
2. Design Process Integration: Facilitates the use of T2I models beyond
mere inspiration, potentially streamlining the concept-to-model
process.
 3. Flexible Design Exploration: Allows designers to adjust the balance
between feasibility and novelty based on the design stage and
requirements.
4. CAD Synergy: Potential for integration with image-to-CAD methods,
enabling seamless transitions between concept generation and
detailed design.
C.7 Limitations and Future Directions
1. Dataset Dependency: The quality of generated designs depends on
the CAD image dataset used.
2. Prompt Balance: Further research on optimizing the balance between
text and CAD image prompts is needed.
3. Validation: Additional case studies across various engineering
domains must validate the method's effectiveness.
4. User Studies: Empirical investigations are needed to understand how
engineers and designers can most effectively use this method in their
workflows.
By addressing these limitations and building upon the strengths of CAD-
Prompted Generative Models, future research can further enhance the
capabilities of T2I models in engineering design, potentially revolutionizing
the concept generation and design refinement processes.

View publication stats