Deep Learning Architectures for AM
CNN, GAN, VAE, Transformer, and Physics-Informed Neural Networks for additive manufacturing applications
Area
Deep Learning + AM
Papers (250+ cit)
2,726
Total Citations
1.8M+
Key Architectures
CNN, GAN, VAE, PINN
AM Tasks
Design, Monitoring, QC
Speedup
100-1000x vs FEM
Deep learning architectures provide the computational backbone for ML applications in additive manufacturing. From CNNs for image-based quality control to GANs for design generation and PINNs for physics-constrained simulation, understanding these architectures is essential for implementing ML in AM workflows.
Overview
Deep learning has revolutionized AM by enabling: (1) real-time process monitoring through image analysis, (2) surrogate models that replace expensive simulations, (3) generative design of novel structures, and (4) physics-constrained predictions that respect conservation laws.
Real-time
Inference Speed
Convolutional Neural Networks
AM Applications
| Application |
Architecture |
Input |
Output |
| Defect classification |
ResNet, VGG |
Layer images |
Defect type |
| Melt pool analysis |
Custom CNN |
Thermal images |
Geometry features |
| Microstructure prediction |
U-Net |
Process parameters |
Grain structure |
| Surface roughness |
CNN regression |
Surface images |
Ra value |
| Topology optimization |
Encoder-decoder |
Boundary conditions |
Material distribution |
Key Architectures
ResNet (Residual Networks)
Skip connections enable training very deep networks. Used for classification tasks with high accuracy.
- Depth: 18, 50, 101, 152 layers
- AM use: Defect classification, quality prediction
U-Net
Encoder-decoder with skip connections. Standard for segmentation and image-to-image translation.
- Strength: Works with small datasets
- AM use: Melt pool segmentation, porosity detection
EfficientNet
Compound scaling of depth, width, resolution. Best accuracy/efficiency tradeoff.
- Versions: B0-B7 (increasing complexity)
- AM use: Real-time monitoring with limited compute
Generative Models (GAN, VAE)
Generative Adversarial Networks
GANs learn to generate realistic samples by adversarial training between generator and discriminator.
| GAN Type |
AM Application |
Key Feature |
| cGAN (Conditional) |
Design generation |
Controls output properties |
| Pix2Pix |
Image-to-image translation |
Thermal to defect maps |
| CycleGAN |
Domain adaptation |
Sim-to-real transfer |
| StyleGAN |
Microstructure synthesis |
High-resolution, controllable |
| 3D-GAN |
Lattice structure design |
Volumetric generation |
Variational Autoencoders
VAEs learn compressed latent representations useful for design exploration:
- Latent space interpolation: Smooth transitions between designs
- Inverse design: Map from properties to designs
- Uncertainty quantification: Probabilistic outputs
- Anomaly detection: High reconstruction error = defect
Diffusion Models
Emerging architecture for high-quality generation:
- Denoising process: Iteratively refines random noise into samples
- AM potential: Design generation, data augmentation
- Current limitation: Slow sampling speed
Physics-Informed Neural Networks
PINNs embed physical laws directly into neural network training, ensuring predictions respect conservation of mass, momentum, and energy.
Formulation
Loss function combines data fit and physics residuals:
- L_data: MSE between predictions and measurements
- L_physics: Residual of governing PDEs (heat equation, Navier-Stokes)
- L_boundary: Boundary condition satisfaction
- Total: L = w1*L_data + w2*L_physics + w3*L_boundary
AM Applications
| Physics |
Governing Equation |
AM Application |
| Heat transfer |
Heat equation |
Melt pool temperature prediction |
| Fluid flow |
Navier-Stokes |
Melt pool convection |
| Solid mechanics |
Elasticity equations |
Residual stress prediction |
| Phase change |
Stefan problem |
Solidification modeling |
Advantages for AM
- Data efficiency: Physics reduces need for labeled data
- Extrapolation: Better generalization outside training domain
- Interpretability: Predictions respect known physics
- Multi-fidelity: Combines simulation and experimental data
Self-Attention Mechanism
Captures long-range dependencies in sequential or spatial data:
- Vision Transformer (ViT): Image patches as tokens
- BERT-style: Pre-training on unlabeled AM data
- Temporal: Process monitoring time series
AM Applications
| Architecture |
Application |
Advantage |
| Vision Transformer |
Layer-wise defect detection |
Global context |
| Temporal Transformer |
Sensor data analysis |
Long-range dependencies |
| Graph Transformer |
Lattice structure analysis |
Topology awareness |
| Point Cloud Transformer |
3D scan processing |
Unordered point sets |
Key Papers
Deep Residual Learning for Image Recognition
He et al., CVPR 2016 | ResNet foundation | 180,000+ citations
Generative Adversarial Nets
Goodfellow et al., NeurIPS 2014 | GAN foundation | 60,000+ citations
Physics-informed neural networks
Raissi et al., JCP 2019 | PINN methodology | 8,000+ citations
Attention Is All You Need
Vaswani et al., NeurIPS 2017 | Transformer architecture | 100,000+ citations
Deep learning for additive manufacturing: State of the art
Survey of DL methods in AM | 1,500+ citations
View all deep learning papers →