Design Optimization References
Primary Source
Paper
Hassan et al. (2024)
Section
8. Design Optimization
Journal
Composites Part C
References
[148]-[160], [124], [139]
Access
Download PDF
This article presents a comprehensive reference list for AI-based design and geometry optimization for 3D printing, extracted from the master review paper by Hassan et al. (2024). The design optimization chapter focuses on integrating AI/ML techniques (VAEs, GANs, CNNs, ANFIS) with topology optimization.
Overview & Benefits
Machine learning-based design optimization for additive manufacturing combines AI techniques (VAEs, GANs, CNNs, reinforcement learning) with topology optimization to automate and enhance the design process. These methods can generate novel geometries, reduce design iteration cycles, and optimize for multiple objectives simultaneously.
30-50%
Design Time Reduction
10-50%
Weight Reduction
6-20%
Cost Decrease
99%
Best Accuracy
About these metrics: These figures represent best-case results from academic studies, primarily on small structural components (brackets, lattice cells, automotive parts). Design time reduction (30-50%) is reported in CNN-based topology optimization studies [157, 158] comparing ML inference time vs. iterative FEA. Weight reduction (10-50%) comes from topology optimization studies [156] on orthotic devices. 99% accuracy refers to material distribution prediction accuracy in [157]. Real-world industrial results may vary significantly depending on part complexity, material, and manufacturing constraints.
Cross-Reference: For research groups working on these topics, see Research Teams & Labs.
Contents
Limitations & Challenges
While ML-based design optimization shows promising results, practitioners should be aware of significant limitations and challenges that are often underreported in the literature.
Manufacturability Constraints
Many ML-generated designs are not directly printable. Common issues include:
- Overhangs: Generative designs often produce unsupported overhangs exceeding 45°, requiring extensive support structures that increase material waste and post-processing time
- Minimum feature size: CNN/GAN outputs may contain features smaller than the printer resolution (typically 0.1-0.4mm for FDM), causing print failures or loss of intended geometry
- Surface finish: Topology-optimized organic shapes often have poor surface quality, requiring CNC post-machining for functional surfaces
- Anisotropy: Most ML models assume isotropic materials, but FDM/SLS parts exhibit significant layer-dependent mechanical properties (up to 50% strength reduction in Z-direction)
- Support removal: Complex internal geometries may trap support material, making removal impossible without destroying the part
Recent work addressing these issues: Liu et al. (2024) on overhang-aware TO, Langelaar (2017) on AM-constrained optimization.
Data & Computational Requirements
- Training data scarcity: Deep learning models require thousands of optimized designs for training. Most organizations lack sufficient validated simulation data, leading to models trained on synthetic or idealized cases
- Simulation-reality gap: Models trained on FEA simulations may not generalize to physical parts due to manufacturing variability, material inconsistencies, and unmodeled physics
- Computational cost: Training 3D generative models (especially for 128³+ voxel resolutions) requires multi-GPU setups and days of training time. Inference is fast, but model development is expensive
- Domain transfer: Models trained on one part family (e.g., brackets) typically don't transfer to other geometries (e.g., heat exchangers) without retraining
Validation & Reliability Concerns
- Limited physical validation: Most papers validate only via simulation; few demonstrate printed parts with measured mechanical properties
- Certification barriers: For safety-critical applications (aerospace, medical), ML-generated designs face regulatory challenges—how do you certify a black-box optimization?
- Reproducibility: Many papers don't release code or trained models, making independent verification difficult
- Overfitting: Models may memorize training geometries rather than learning generalizable optimization principles
Common Workflows & When to Use What
Different ML approaches suit different design scenarios. This section provides guidance on selecting appropriate methods.
Workflow Comparison
| Approach | Best For | Limitations | Data Needs | Maturity |
|---|---|---|---|---|
| Traditional TO + FEA | Single-objective optimization, well-defined load cases | Slow iteration, no aesthetics, manufacturability issues | Low (physics-based) | Industrial |
| CNN/U-Net Surrogates | Rapid TO inference after training, design space exploration | Requires large training sets, limited to trained geometry classes | High (1000s of TO solutions) | Research |
| VAE/GAN Generative | Novel geometry generation, aesthetic + performance balance | Mode collapse, manufacturability not guaranteed | High (diverse design corpus) | Research |
| Reinforcement Learning | Sequential decision problems, process parameter optimization | Sample inefficient, reward engineering difficult | Medium (environment simulation) | Emerging |
| Neural Field/Implicit | Continuous representations, multi-parameter optimization | Slow training, complex implementation | Medium | Emerging |
Decision Guide
Choose your approach based on these factors:
- If you need certified, predictable results: Use traditional TO with commercial software (ANSYS, Altair, nTopology). ML is not yet mature enough for safety-critical applications without extensive validation
- If you're exploring many design variants quickly: Train a CNN surrogate on your geometry class, then use it for rapid what-if analysis
- If aesthetics matter as much as performance: Use VAE/GAN approaches (Oh et al. 2019) that can balance multiple objectives including visual appeal
- If you need manufacturable designs: Use manufacturability-constrained methods (Williams et al. 2023) or add post-processing checks; don't trust raw ML outputs
- If you're optimizing process parameters (not geometry): Consider reinforcement learning for print path, layer thickness, or temperature optimization
- If data is scarce: Stick with physics-based TO; ML methods need substantial training data to outperform traditional approaches
Typical Hybrid Workflow
A practical production workflow often combines multiple methods:
- Requirements → ML-assisted concept generation (GAN/VAE for initial shapes)
- Concept → Topology optimization (traditional SIMP/BESO for structural refinement)
- TO result → Manufacturability check (rule-based or ML classifier for printability)
- Geometry → Simulation validation (FEA for structural, CFD for thermal)
- Validated design → Process optimization (ML for print parameters, orientation)
- Print → Digital twin feedback (ML-based process monitoring, defect detection)
- Physical testing → Model update (close the loop with real data)
Most Influential Papers
Ranked by citation count as of December 2024.
Most Cited
2019
[153] Deep Generative Design: Integration of Topology Optimization and Generative Models
~263 citations
- Method: VAEs + GANs (BEGANs)
- Key Contribution: First framework integrating generative models with topology optimization
- Innovation: Considers both engineering performance AND aesthetics
- DOI: 10.1115/1.4044229
- PDF: Download Full Paper (2.4MB)
2017
[154] A Hybrid Machine Learning Approach for AM Design Feature Recommendation
~50+ citations
- Method: SVM classifier + hierarchical clustering
- Application: Racing car components
- DOI: 10.1108/RPJ-03-2016-0041
2016
[155] A Cost-Driven Design Methodology for AM Variable Platforms
~31 citations
- Method: Multi-objective GA + Mamdani fuzzy inference
- Goal: Minimize cost while maximizing process flexibility
- DOI: 10.1115/1.4032504
Core Design Optimization Papers
| Ref # | Authors | Title | Journal | Year | Citations |
|---|---|---|---|---|---|
| [153] | Oh, Jung, Kim, Lee, Kang | Deep generative design: Integration of topology optimization and generative models | J. Mech. Des. | 2019 | ~263 |
| [154] | Yao, Moon, Bi | A hybrid machine learning approach for AM design feature recommendation | Rapid Prototyping J. | 2017 | ~50+ |
| [155] | Yao, Moon, Bi | A cost-driven design methodology for AM variable platforms | J. Mech. Des. | 2016 | ~31 |
| [156] | Kumar, Chhabra | Parametric topology optimization for customized orthotic appliances | Mech. Adv. Mater. Struct. | 2023 | <10 |
| [157] | Rade, Jignasu, Herron et al. | Deep learning-based 3D multigrid topology optimization | Eng. Appl. Artif. Intell. | 2023 | <10 |
| [158] | Rasulzade et al. | Reduction of material usage using U-net CNN topology optimization | Eurasian Chem. J. | 2022 | <10 |
| [159] | Saleh et al. | Prediction of mechanical properties for CF/PLA lattice structures | Polymers | 2023 | <10 |
| [160] | Pollak, Torok | Use of generative design tools in 3D printing | TEM Journal | 2022 | <10 |
| [139] | Grozav et al. | ANN-based predictive model for FEA of AM components | Machines | 2023 | <10 |
| [124] | Fouly et al. | Mechanical properties of annealed 3D-printed PLA-date pits composite | Polymers | 2023 | <10 |
Review Papers
| Ref # | Authors | Title | Journal | Year |
|---|---|---|---|---|
| [148] | Mountstephens, Teo | Progress and challenges in generative product design: a review | Computers | 2020 |
| [149] | Barbieri, Muzzupappa | Performance-driven engineering design based on generative design and TO | Applied Sciences | 2022 |
| [150] | Brossard et al. | How Generative Design Could Reshape Product Development | McKinsey & Co | 2020 |
Papers by Methodology
1. Generative Models (VAE/GAN)
VAEs + GANs
[153] Oh et al. (2019)
- First framework integrating generative models with topology optimization
- Uses BEGANs (Boundary Equilibrium GANs)
- Anomaly detection for design novelty assessment
- Case study: 2D wheel design problem
2. Topology Optimization with Deep Learning
Pyramid U-Net CNN
99% Accuracy
[157] Rade et al. (2023) & [158] Rasulzade et al. (2022)
- Multigrid techniques + distributed GPU training
- High-resolution 3D geometries (128x128x128)
- 4.77x training speedup
- 99% material distribution accuracy
ANSYS Workbench TO
42.6% Weight Reduction
[156] Kumar & Chhabra (2023)
- 3D scanning-assisted parametric modeling
- 149.19% better heat dissipation
- Application: Custom medical splints via FDM
3. Machine Learning for Design Recommendation
SVM + Clustering
[154] Yao, Moon, Bi (2017)
- SVM classifier + progressive dendrogram cutting
- Application: Racing car components (suspension, bumper, driveshaft)
- Reduces labor and time consumption for designers
4. TPMS/Lattice Structure Optimization
ANFIS Model
[159] Saleh et al. (2023)
- Structures: Diamond, Gyroid, Primitive cell topologies
- Materials: PLA and carbon fiber-reinforced PLA
- Max 7.61% deviation (vs 21.11% for mathematical models)
5. Predictive Models for FEA
ANN + Adamax
93% Accuracy
[139] Grozav et al. (2023)
- 3 hidden layers with 8 nodes each
- Application: Orthotropic material profile for FDM-printed PLA
ANFIS (5 layers)
<0.2% Error
[124] Fouly et al. (2023)
- Hardness: 9.88x10-3% error
- Modulus: 0.18% error, Strength: 0.08% error
- Application: Orthotic biomedical applications
Latest Research (2024-2025)
This section highlights the most significant recent advances in ML-based design optimization for additive manufacturing, verified through peer-reviewed publications from 2024-2025.
1. Physics-Informed Neural Networks for Topology Optimization
Physics-informed neural networks (PINNs) have emerged as a transformative approach, eliminating the need for FEA during optimization by embedding physical laws directly into the network architecture.
Key Paper
2025
CPINNTO: Complete Physics-Informed Neural Network for Topology Optimization
Jeong et al. — Engineering Structures, January 2025
- Key Innovation: Uses two PINNs to replace both structural and sensitivity analyses—no labeled data or FEA required for training
- Applications: Periodic, multi-scale, multi-material, and geometrically nonlinear topology optimization
- Significance: First complete PINN framework applicable to complex industrial TO challenges
PINN + Deep Energy
2025
FF-PINNTO: Fourier Feature-Embedded PINN for Nonlinear Structures
Computer Methods in Applied Mechanics and Engineering, 2025
- Integrates Deep Energy Method with neural reparameterization
- Handles geometrically nonlinear (hyperelastic) structures
- Mesh-free approach for complex boundary conditions
Discrete PINN
2025
dPINN: Discrete Physics-Informed Neural Networks for Large-Scale TO
Computer Methods in Applied Mechanics and Engineering, 2025
- Addresses memory limitations for large-scale geometrically nonlinear TO
- Mesh-based local interpolation reduces computational demands
- Enhanced resilience against mesh distortion compared to traditional FEM
2. Generative Models with Manufacturability Constraints
Key Paper
2024
Pix2Pix-GAN for AM-Constrained Topology Optimization
Journal of Manufacturing and Materials Processing, October 2024
- Key Innovation: Conditional GAN trained on color-coded images for cantilever optimization
- Result: Eliminates iterative TO steps, reducing design cycle from hours to minutes
- AM Integration: Considers additive manufacturing restrictions during training
RL + PPO
2025
Reinforcement Learning-Based Topology Optimization for Lightweight Structures
- Integrates TO with deep reinforcement learning (Proximal Policy Optimization)
- Adheres to strict constraints: Von Mises stress ≤300 MPa, displacement ≤0.5mm
- SDF smoothing + STL export for direct 3D printing
- Result: Up to 40% weight reduction while maintaining compliance
Diffusion + GAN
2024
Diffusion-GAN for Structural Topology Optimization
Engineering Applications of AI, 2024
- Adaptive diffusion process resolves GAN training instability
- Guarantees precision and high quality of generated topology results
- Emerging approach combining diffusion model stability with GAN generation speed
3. Fiber-Reinforced Composite Optimization
Key Paper
2025
Neural Co-Optimization of Topology, Layers, and Fiber Orientation
ACM Transactions on Graphics (SIGGRAPH), 2025
- Key Innovation: Three implicit neural fields simultaneously optimize shape, layer sequence, and fiber orientation
- Manufacturability: Directly formulates AM constraints into differentiable optimization
- Result: 33.1% improvement in failure loads for fiber-reinforced thermoplastic composites
AI Multiphysics
2025
AI-Driven Multiphysics for Fiber Orientation Distribution Prediction
Materials Today Communications, 2025
- Genetic algorithm + ANN hybrid for fiber orientation distribution (FOD) prediction
- RSC model coefficient optimization for both thermoplastic and thermosetting polymers
- Applicable to FDM/FFF continuous fiber composites
FRC-TOuNN
2024
FRC-TOuNN: Topology Optimization of Continuous Fiber Composites
- Concurrent optimization of matrix topology and fiber distribution
- Smooth fiber paths for manufacturing compatibility and numerical stability
- Functionally graded continuous fiber-reinforced composites
4. Digital Twin Integration & Real-Time Control
Key Paper
2024
Digital Twin Framework with LSTM + Bayesian Optimization for DED
Journal of Manufacturing Processes, 2024 | arXiv
- Process: Laser directed-energy deposition (DED)
- Method: LSTM-based ML with Bayesian Inference for real-time temperature prediction
- Innovation: Unified framework integrating real-time monitoring, physics-based modeling, and control
TiDE + MPC
2025
Real-Time Decision-Making with Model Predictive Control
Journal of Manufacturing Systems, 2025
- Time-Series Dense Encoder (TiDE) as surrogate model for MPC
- Simultaneous multi-step prediction for proactive control
- Enables autonomous manufacturing with real-time quality adjustment
Unity + ML
2023
Integrated ML-Digital Twin for FDM with Real-Time Defect Detection
- Unity + OctoPrint + Raspberry Pi integration
- 92% Average Precision for defect detection (91% defected, 94% non-defected)
- Real-time monitoring and control pipeline
5. Reinforcement Learning for Process Optimization
RL + Porosity
2024
RL-Based Porosity Prediction for L-PBF Parameter Optimization
Scripta Materialia, September 2024
- First RL approach for porosity prediction in metal laser-powder bed fusion
- State space: permutations of laser power, scan speed, hatch spacing
- Experimentally validated on high-strength A205 Al alloy
Q-Learning
2024
Optimal Data-Driven Control for Wire Arc AM
Journal of Intelligent Manufacturing, January 2024
- Off-policy RL (Q-learning) for laser power + scan velocity optimization
- Objective: maintain steady-state melt pool depth
- WAAM simulator for realistic training environment
Vision + RL
2024
Vision-Based Uncertainty Quantification with RL for Extrusion Control
arXiv / Additive Manufacturing, 2024
- Agent dynamically adjusts flow rate and temperature in real time
- Vision-based uncertainty quantification module
- Addresses training efficiency and uncertainty management bottlenecks
6. Recent Review Papers
| Title | Venue | Date | Focus |
|---|---|---|---|
| Machine Learning in AM: Enhancing Design, Manufacturing and Performance | J. Intelligent Manufacturing | Jan 2025 | Comprehensive review including 4D printing + ML |
| Review of ML Applications in Additive Manufacturing | Results in Engineering | Dec 2024 | Quality, sustainability, defect detection |
| Current Applications of ML in AM: Challenges and Future Trends | Archives of Computational Methods | Dec 2024 | Fourier neural operators, dynamic PINNs for TO |
| Progress and Opportunities for ML in Materials and Processes of AM | Advanced Materials (Wiley) | Mar 2024 | Supervised, semi-supervised, RL, transformers |
| Big Data, ML, and Digital Twin Assisted AM: A Review | Materials & Design | Jun 2024 | Data integration, digital twin frameworks |
| Recent Advances in Design Optimization and AM of Composites | npj Advanced Manufacturing | 2025 | Nanoparticle, short fiber, continuous fiber composites |
Key Trends Summary
- PINNs replacing FEA: Physics-informed neural networks now handle complex nonlinear TO without labeled data or traditional simulation
- Manufacturability by design: New frameworks (RL+SDF, constraint-aware VAEs) ensure printability from the start, not as post-processing
- Composite co-optimization: Simultaneous optimization of topology + fiber orientation + layer paths is now achievable via neural implicit fields
- Real-time digital twins: LSTM, TiDE, and MPC enable closed-loop control for DED, L-PBF, and WAAM processes
- RL for process control: Q-learning and PPO optimize parameters (power, speed, temperature) for quality outcomes rather than just geometry
- Diffusion models emerging: Diffusion-GAN hybrids promise more stable training for generative design
Practical Guidance for Engineers
This section provides actionable recommendations for engineers and designers looking to implement ML-based design optimization in their workflows.
Data Requirements
What data do you need to train ML design models?
- For CNN/U-Net surrogates: 1,000-10,000 topology optimization solutions (boundary conditions → optimized geometry pairs). Generate via traditional TO tools (ANSYS, TopOpt, OpenTOP)
- For VAE/GAN generative models: Diverse corpus of validated designs with performance labels. Quality matters more than quantity—100 well-characterized designs may beat 10,000 synthetic ones
- For process optimization: Print job logs with parameters (speed, temp, layer height) and outcome metrics (dimensional accuracy, strength, surface roughness)
- For digital twins: Real-time sensor data (thermal cameras, acoustic emission, layer images) paired with defect labels
Data quality checklist:
- ☐ Consistent mesh resolution across all samples
- ☐ Validated FEA results (mesh convergence study)
- ☐ Realistic boundary conditions (not just textbook cases)
- ☐ Material properties matched to your actual printing materials
- ☐ At least some physical validation (printed + tested parts)
Tools & Software
| Category | Tool | Description | Access |
|---|---|---|---|
| Topology Optimization | ANSYS Mechanical, Altair OptiStruct | Industrial TO with AM constraints | Commercial |
| nTopology | Lattice + TPMS design, AM-focused | Commercial | |
| TopOpt (DTU) | Open MATLAB/Python TO code for research | Open source | |
| ML Frameworks | PyTorch, TensorFlow | General ML for custom models | Open source |
| DeepXDE | Physics-informed neural networks | Open source | |
| AM-specific ML | 3DXpert, Materialise Magics | Build preparation with some ML features | Commercial |
| Research Repos | Open3D-ML | 3D ML library with point cloud/mesh support | Open source |
Common Pitfalls to Avoid
- Garbage in, garbage out: If your training data has non-manufacturable designs, your model will generate non-manufacturable designs. Curate your training set carefully
- Overfitting to benchmarks: A model that achieves 99% accuracy on a benchmark (e.g., 2D cantilever beam) may fail completely on real industrial geometries. Always test on held-out realistic cases
- Ignoring material anisotropy: FDM/FFF parts have ~50% reduced strength in the Z-direction. If your training data assumes isotropic material, your optimized designs may fail structurally
- Skipping physical validation: Simulation-optimal ≠ real-world optimal. Budget time and parts for physical testing before production deployment
- Over-automation: Keep engineers in the loop. ML should augment expertise, not replace critical thinking about manufacturability and safety
- Underestimating post-processing: Many ML-generated designs require significant CAD repair (smoothing, feature removal, surface reconstruction) before they're printable. Factor this into your timeline
Getting Started: Recommended Path
- Week 1-2: Establish baseline with traditional TO (use TopOpt MATLAB code or commercial tools). Document performance metrics on your target geometry class
- Week 3-4: Generate training data—run 500-1000 TO solutions with varying loads, constraints, volume fractions
- Week 5-6: Train a simple CNN (ResNet or U-Net) to predict TO results. Compare inference time vs. traditional TO
- Week 7-8: Add manufacturability constraints. Either filter outputs or retrain with AM-constrained data
- Week 9+: Physical validation—print and test ML-generated vs. traditional TO designs. Iterate based on findings
Note: These timelines assume dedicated ML and CAD engineering resources. Academic teams should expect 3-6 months for a publishable result; industrial deployment may take 6-12 months.
DOI Links
| Ref # | Citations | DOI |
|---|---|---|
| [153] | ~263 | doi.org/10.1115/1.4044229 |
| [154] | ~50+ | doi.org/10.1108/RPJ-03-2016-0041 |
| [155] | ~31 | doi.org/10.1115/1.4032504 |
| [156] | <10 | doi.org/10.1080/15376494.2023.2214908 |
| [157] | <10 | doi.org/10.1016/j.engappai.2023.107033 |
| [158] | <10 | doi.org/10.18321/ectj1471 |
| [159] | <10 | doi.org/10.3390/polym15071720 |
| [160] | <10 | doi.org/10.18421/TEM111-31 |
| [139] | <10 | doi.org/10.3390/machines11050547 |
| [124] | <10 | doi.org/10.3390/polym15163395 |
See also
- Research Teams & Labs
- Top Journals
- Machine Learning for Additive Manufacturing (main topic page)
References and Source Documents
Primary Source
A Review of AI for Optimization of 3D Printing of Sustainable Polymers and Composites
Affiliation: University of Guelph, Canada
- Journal: Composites Part C: Open Access, Volume 15, 2024, Article 100513
- Published: 27 September 2024
- Section Used: Section 8 — AI-Based Design and Geometry Optimization for 3D Printing
- Total References: 184 (this page covers refs [124], [139], [148]-[160])
Access the Original:
- DOI: 10.1016/j.jcomc.2024.100513 (Publisher - Open Access)
- ScienceDirect Full Text
- Download PDF (Local Copy)
Downloaded Papers
| Ref | Paper | Access |
|---|---|---|
| [153] | Oh et al. (2019) — Deep Generative Design | PDF | arXiv | DOI |
| [154] | Yao, Moon, Bi (2017) — Hybrid ML for AM Design | ResearchGate | DOI |
| [155] | Yao, Moon, Bi (2016) — Cost-Driven Design | ResearchGate | DOI |
External Databases
- Semantic Scholar — Citation tracking and related papers
- Google Scholar — Papers citing this review
- Composites Part C: Open Access — Journal homepage
How to Cite
Hassan, M., Misra, M., Taylor, G.W., & Mohanty, A.K. (2024). A review of AI for optimization of 3D printing of sustainable polymers and composites. Composites Part C: Open Access, 15, 100513. https://doi.org/10.1016/j.jcomc.2024.100513