PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep Generative Models in Engineering Design: A Review

Lyle Regenwetter, Amin Heyrani Nobari, Faez Ahmed

arXiv:2110.10863v4cs.LGstat.ML

TL;DR

Engineering design needs automated synthesis that can improve productivity, customization, and access while reducing dependence on manual expertise. This review analyzes how Deep Generative Models are used across engineering design through their algorithms, representations, applications, datasets, and auxiliary methods. It finds substantial progress across design domains, alongside persistent challenges involving sparse data, performance evaluation, creativity, constraints, feasibility, and manufacturability.

  • Problem

    Engineering design automation seeks to reduce costly dependence on human expertise while enabling customized and optimized products, but adapting generative machine learning to engineering remains challenging.

  • Method

    The paper reviews DGM algorithms, design representations, engineering applications, datasets, and auxiliary methods, including 63 papers from multiple design subdisciplines.

  • Results

    DGMs have been applied across structural optimization, materials design, shape synthesis, topology optimization, and other engineering-design tasks, with methods addressing performance and high-resolution synthesis.

  • Takeaways & Limitations

    The review provides a reference for selecting methods and representations and identifies solution pathways for advancing data-driven generative design research.

  • Takeaways & Limitations

    Engineering DGMs remain constrained by sparse and uneven data, brittle surrogate evaluation, invalid generated designs, and unresolved feasibility and manufacturability requirements.

Abstract

from arXiv · show

Automated design synthesis has the potential to revolutionize the modern engineering design process and improve access to highly optimized and customized products across countless industries. Successfully adapting generative Machine Learning to design engineering may enable such automated design synthesis and is a research subject of great importance. We present a review and analysis of Deep Generative Machine Learning models in engineering design. Deep Generative Models (DGMs) typically leverage deep networks to learn from an input dataset and synthesize new designs. Recently, DGMs such as feedforward Neural Networks (NNs), Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), and certain Deep Reinforcement Learning (DRL) frameworks have shown promising results in design applications like structural optimization, materials design, and shape synthesis. The prevalence of DGMs in engineering design has skyrocketed since 2016. Anticipating continued growth, we conduct a review of recent advances to benefit researchers interested in DGMs for design. We structure our review as an exposition of the algorithms, datasets, representation methods, and applications commonly used in the current literature. In particular, we discuss key works that have introduced new techniques and methods in DGMs, successfully applied DGMs to a design-related domain, or directly supported the development of DGMs through datasets or auxiliary methods. We further identify key challenges and limitations currently seen in DGMs across design fields, such as design creativity, handling constraints and objectives, and modeling both form and functional performance simultaneously. In our discussion, we identify possible solution pathways as key areas on which to target future work.

1 Introduction

Generative design automates design synthesis by learning from existing information, offering potential gains in productivity, customization, and access. This review organizes DGM methods, representations, applications, datasets, and challenges while emphasizing that engineering applications remain immature.

  • Motivation and background: Automating design could reduce costs, increase industrial productivity, and support customized products without prohibitive manual-design expense.The motivation is linked to broader potential benefits for global productivity, prosperity, equity, and access.
  • Motivation and background: Generative design algorithms directly synthesize designs through explicit programming or implicit learning from existing information.Deep Generative Models approximate complex, high-dimensional distributions and generate new samples using deep learning.
  • Challenges: Engineering design DGMs face challenges in performance-aware inverse design, sparse data, creativity, and physical feasibility or manufacturability.Performance evaluation is difficult because numerical simulation is costly and real-world evaluation can be even harder.
  • Challenges: Researchers have proposed methods incorporating performance, optimization, novelty, and creativity, but engineering-design DGMs remain in their infancy.The review identifies these challenges as requiring further effort rather than considering them resolved.
  • Review scope: The review surveys common DGM methods, design parameterizations, applications, datasets, and field-wide challenges as a reference for researchers.It also presents Figure 1 as an overview of the standard engineering-design application process.

2 Overview of Deep Generative Models

The review distinguishes direct, adversarial, latent-space, and sequential approaches to design generation, with GANs, VAEs, and RL most common. It also describes conditioning, training difficulties, and the trade-offs between data dependence, reward design, and objective flexibility.

  • Model families: Four common DGM approaches generate designs directly with DNNs, adversarially with GANs, from embeddings with VAEs, or sequentially with RL.GANs, VAEs, and RL are observed to be the most commonly used approaches in design synthesis.
  • Generative Adversarial Networks: GANs use a generator to map noise to designs while a discriminator learns to distinguish generated samples from real data.The two models are trained adversarially to support synthesis from the design distribution.
  • Generative Adversarial Networks: GAN training can be unstable when the discriminator overpowers the generator, causing vanishing gradients and halted generator training.The WGAN modifies the discriminator, loss function, and training process to estimate Wasserstein distance.
  • Generative Adversarial Networks: GAN mode collapse occurs when generated samples fail to cover the data distribution or consist of only a few discriminator-fooling designs.Researchers have developed techniques intended to reward diversity in generated samples.
  • Conditional generation: Conditioning imposes design constraints, requirements, preferences, or continuous performance targets through mechanisms such as cGANs, InfoGANs, and conditional VAEs.Conditional VAEs add a conditioning vector to both encoder and decoder, while GAN conditioning can involve the generator and discriminator.
  • Variational Autoencoders: VAEs encode designs into a structured probabilistic latent space and decode sampled latent vectors into new designs.The latent distribution is regularized through probabilistic sampling and a KL-divergence loss.
  • Reinforcement Learning: RL constructs designs through sequential actions and reward signals, requiring meaningful evaluation but allowing nondifferentiable objectives.Unlike GANs and VAEs, RL does not require a dataset, although it may depend on high-fidelity simulation environments.

3 Overview of Design Representation Methods

Engineering DGMs use representations including images, voxels, point clouds, meshes, SDFs, parametric data, grammars, and graphs. Representation choice affects visualization, simulation, fabrication, interpretability, flexibility, and the suitable generative architecture.

  • Representation choice: Representation methods determine how designs are encoded and influence the DGM architectures suitable for particular synthesis problems.The review compares definitions, strengths, and weaknesses of common methods.
  • Grid-based representations: Images and voxels provide grid-based representations compatible with established computer-vision methods but can be difficult to fabricate or simulate directly.Image-based workflows may require conversion to 3D models before conventional FEA or CFD evaluation.
  • Geometric representations: Point clouds represent complex geometry with finite point sets and are often easy to obtain from 3D scanning, but commonly require conversion to meshes or BRep.Their fidelity can vary with the point representation.
  • Geometric representations: Meshes are common 3D object representations that can be directly visualized and simulated in many FEA and CFD tools.Triangular meshes are widely used in computer graphics and finite-element software.
  • Geometric representations: SDFs encode signed distances from coordinate points to object surfaces, with sign indicating whether points lie inside or outside.They can be represented as rasterized grids containing continuous voxel values.
  • Parametric representations: Parametric representations organize design parameters into collections that may support fabrication, human interaction, and conditioning on high-level goals.Parameters can sometimes be linked directly to a generative model’s latent variables.
  • Symbolic and relational representations: Grammars encode designs through variables, symbols, and expansion rules, while graphs encode systems as nodes and edges.Grammars can dictate feasible assembly hierarchies, whereas graphs flexibly describe complex systems.
  • Symbolic and relational representations: Graph-based DGMs remain less common in engineering design, possibly because graph-based design datasets are scarce.This representation has nevertheless supported design generation and optimization in other contexts, including molecular graph generation.

4 Literature Review Methodology

The review selects engineering-design papers involving deep generative models, emphasizing VAEs, GANs, and reinforcement learning. Its search covered specified design venues and papers published from January 2014 through September 2021.

  • Selection criteria: Included studies had to use deep learning and DGMs, particularly VAEs, GANs, or reinforcement learning, while excluding works focused only on design optimization.The scope was restricted to engineering design rather than other domains.
  • Selection criteria: The review considered papers published between Jan. 2014 and Sep. 2021.This date boundary formed part of the stated selection criteria.
  • Search procedure: The search used Google Scholar queries for “Generative Adversarial Network,” “Variational Autoencoder,” and “Reinforcement Learning.”It initially focused on the Journal of Mechanical Design, IDETC proceedings, Computer-Aided Design Journal, ICED, and AIEDAM.

5 Application Domains in Engineering Design

The review organizes engineering-design applications of DGMs across topology optimization, materials and metamaterials, airfoils, 3D shapes, and mechanisms, using multiple architectures and data representations.

  • Topology Optimization: Topology-optimization studies use DGMs to predict optimized topologies, initialize conventional optimization, iteratively explore design-space gaps, incorporate human selection, and generate high-resolution results.Supervised generation targets computational savings, while iterative and super-resolution methods address exploration and resolution-cost challenges.
  • Materials and Metamaterials: Materials-design applications use DGMs to generate microstructures, metamaterial unit cells, and designs targeting material properties through inverse-design workflows.These applications may combine generative models with latent-space optimization to attain desired macroscopic properties.
  • Airfoil Synthesis: Airfoil studies apply DGMs to two-dimensional geometric synthesis while addressing performance and novelty, including PaDGAN generation evaluated with lift-to-drag ratio.PaDGAN uses a diversity mechanism to generate high-quality, previously unseen airfoil designs and has been extended to multi-objective generation.
  • 3D Shape Synthesis: Three-dimensional shape studies use voxels, point clouds, or meshes, with engineering applications emphasizing functional performance or manufacturability beyond visual appearance.Examples include aircraft optimization with aerodynamic drag objectives and CFD-based evaluation.
  • Other Applications: Kinematic-synthesis studies use VAEs and reinforcement learning to generate mechanisms for motion curves or power transmission, sometimes enabling human customization or straight-line generation.VAE methods expose latent-space perturbations for designer interaction, while Deep Q-Learning synthesizes planar linkages for straight-line motion.

6 Datasets

The review surveys datasets used or potentially useful for data-driven design and emphasizes the need for larger, well-annotated public collections. Topology-optimization studies commonly generate their own datasets, many of which are not publicly available.

  • Dataset Overview: The dataset survey covers commonly used or potentially useful collections for training DGMs in engineering design, while noting that the list is not comprehensive.The authors hope researchers will create larger, well-annotated datasets and release them publicly.
  • Topology-Optimization Datasets: Most topology-optimization papers generate their own datasets, which are often unavailable publicly.SIMP is the most common dataset-generation method, although some open-source topology datasets exist.
  • Topology-Optimization Datasets: One open dataset contains 10,000 artificially generated topologies, including final and intermediate 40x40 topology images across 100 optimization steps.The dataset provides topology states throughout the optimization process rather than only final designs.

6.2 Microstructure Datasets

Microstructure datasets commonly derive from microscopy and materials databases, supporting generative modeling and materials-design tasks. Public resources include scan-image collections, composite-material databases, and synthetic microstructure generators.

  • Microstructure Data: Optical and scanning electron microscopy have produced numerous publicly available datasets of material microstructure scan images.The review identifies OM and SEM as established technologies for visualizing microstructures.
  • Materials Databases: The NanoMine nanopolymer composite database contains over 20,000 data points for composite-material design.The review also notes compiled lists of materials-science datasets and synthetic microstructure datasets.
  • Synthetic Microstructures: Yang et al. provide a trained GAN model that generates synthetic microstructure images.Synthetic data generation complements publicly available microscopy and materials datasets.
  • Related Shape Dataset: The UIUC airfoil database contains nearly 1,600 real-world airfoil designs represented by surface coordinates.Researchers standardize inconsistent coordinate counts using B-spline interpolation.

6.4 3D Object Datasets

The review describes widely used 3D-object datasets spanning general shapes and mechanical components, with representations including voxels, point clouds, meshes, and 2D views.

  • General 3D Models: ShapeNet contains over 51,300 3D models across 55 object categories, while ModelNet contains 127,915 voxel-based models across 662 categories.Both datasets are used by multiple engineering-design studies discussed in the review.
  • General 3D Models: PartNet extends ShapeNet with fine-grained hierarchical semantic annotations for object components.Its annotations support component-level organization beyond whole-object categories.
  • Engineering Applications: ShapeNet and ModelNet are used in engineering-design studies, including work on aircraft and other three-dimensional models.The review identifies both as commonly used 3D model datasets in its application survey.
  • Mechanical Components: The Mechanical Components Benchmark contains 58,696 models from 68 classes with hierarchical labels and representations as point clouds, voxels, and 2D views.Its label tree groups components into progressively specific subclasses, such as fasteners and nuts.

6.5 CAD and CAD-based Datasets

The review highlights two CAD datasets that support generative design through diverse model representations and parameterizations.

  • 8,625 Autodesk Fusion models form a reconstruction dataset containing sequential CAD operation steps for part generation.
  • 35,680 Autodesk Fusion models form a segmentation dataset for CAD-related learning tasks.
  • BIKED contains 4,512 bicycle CAD models with assembly images, segmented subcomponent images, and 2,395 mixed-type design parameters.

6.6 Metamaterials Datasets

The reviewed metamaterials datasets pair unit-cell representations with stiffness or elastic tensor data, supporting generative design and topology optimization research.

  • 248,396 2D unit cells are represented by 50x50 pixelated matrices with associated stiffness tensor components.The dataset can also support topology optimization research.
  • 3,000 3D isosurface unit cells are sampled from 30 level-set functions and paired with corresponding 3D elastic tensor components.
  • 795 unit cells are generated from 10 lattice models with associated stiffness tensors.

6.7 Sketch Datasets

The review describes sketch datasets spanning broad user-generated categories, innovative product concepts, and recorded human design operations with performance measures.

  • QuickDraw provides 50 million user-drawn doodles across 345 categories for sketch-based generative modeling.The data was collected by Google through an interactive sketching game.
  • A milk-frother dataset contains 934 innovative design sketches paired with text descriptions.The associated descriptions could support generative models incorporating natural-language processing.
  • A truss dataset records sequential design operations, geometric parameters, safety factors, and weights from sixteen human teams.It is primarily intended to study or mimic the human design process.

7 Discussion, Challenges and Future Work

The review identifies performance evaluation, feasibility, creativity, benchmarking, and data sparsity as central challenges for engineering-design DGMs, while outlining promising research directions.

  • 7.1 Design Performance Evaluation: Performance-aware DGMs face fidelity, cost, and differentiability challenges when evaluating generated designs.Physical tests offer high fidelity but are costly; simulations offer medium fidelity but may also be costly and nondifferentiable.
  • 7.1 Design Performance Evaluation: Surrogate models are the most common low-cost evaluation method, but may be brittle and generalize poorly beyond their training data.Self-supervised augmentation and multi-fidelity modeling are proposed to improve surrogate accuracy.
  • 7.2 Feasibility, Constraints, and Manufacturability: Explicit and implicit design constraints, including physical feasibility and manufacturability, remain difficult for probabilistic DGMs to satisfy reliably.Inexpensive, reliable validation methods are a potential solution but may require significant human input.
  • 7.2 Feasibility, Constraints, and Manufacturability: Design representations can create a domain gap because generated images may not translate to practical forms and some 3D models require additive manufacturing.Manufacturability is modeled in relatively few reviewed works, while parametric manufacturing data is challenging to learn and generate.
  • 7.3 Creativity and Novelty: DGMs commonly mimic existing designs, which supports realism and sample quality but discourages creative or novel outputs.Recent methods such as CreativeGAN introduce novel features into typical designs to expand generated design-space novelty.
  • 7.4 Evaluating Model Performance & Benchmark Problems: State-of-the-art comparisons are application-dependent because studies often use similar methods on specific datasets with few baselines.The review calls for broadly applicable benchmarks, larger datasets, and higher-quality evaluation resources.
  • 7.4 Evaluating Model Performance & Benchmark Problems: Benchmarking is difficult because datasets are small, domain-restricted, sparse, often proprietary, and paired with incompatible representations.The lack of benchmark problems and performance metrics further limits comparisons across approaches.
  • 7.5 Data Limitations and Quality: Data sparsity reflects limited public coverage, insufficient dataset size for data-hungry models, and underrepresented design representations such as graphs.The review encourages large public datasets and effective data augmentation methods.

8 Conclusion

The review surveys DGMs for engineering design, covering their algorithms, representations, applications, datasets, and current challenges. It also identifies possible solution pathways for these challenges.

  • The review overviews DGMs commonly used for engineering design, emphasizing GANs, VAEs, and reinforcement learning.
  • It compares the strengths and weaknesses of common design representation and parameterization methods.
  • The authors collect and review 63 papers that directly propose DGMs across engineering design subdisciplines.
  • The review catalogs commonly used datasets for developing and testing data-driven design methods.
  • It highlights field-wide challenges and limitations while presenting possible solution pathways for future work.
Loading 2110.10863v4…