Core Contributions and Methodology

Training Strategy

To handle large build-plate images:

• Patch size: 64×64
  
• 115 patches per layer
  
• 80:20 train-test split
  
• Autoencoder training: 100 epochs
  
• Diffusion training: 300 epochs

A second experiment performs part-based training, ensuring no part-level data leakage.

Evaluation Metrics

3D Morphology Reconstruction

To evaluate geometric fidelity beyond 2D metrics:

1. Each layer is segmented using Segment Anything (SAM).
2. Masks are stacked to reconstruct 3D morphology.
3. Three geometry metrics are computed:

• IoU (Intersection-over-Union)
• Hausdorff Distance
• Voxel mismatch

Latent diffusion improves IoU and reduces geometric error compared to low-resolution and bicubic interpolation baselines.

Surface Roughness Estimation

Surface roughness metrics are computed from reconstructed contours:

• Ra (Mean absolute roughness)
• Rq (Root-mean-square roughness)
• Rz (Peak-to-valley height)

Example (Dataset A):

• HR: Ra = 30.7 µm
• Latent Diffusion: Ra = 32.8 µm
• Bicubic Upsampling: Ra = 57.8 µm

Latent diffusion closely approximates true surface roughness.

Zero-Shot Generalization

The study evaluates transferability across part geometries:

• Synthetic low-resolution images created via Gaussian degradation.
• Entire part builds held out during training.
• Model tested on unseen geometries.

Result:

• Stable PSNR/SSIM performance
• Covariance distance improves as training diversity increases
• Robust performance under increasing blur kernel size

This demonstrates strong inter-layer and inter-part generalization.

Key Findings

• Latent diffusion reconstructs fine powder-bed texture.
• Significant reduction in MAE and covariance distance.
• Improved PSNR and SSIM.
• ~13× faster inference than pixel diffusion.
• Improved 3D geometric accuracy (IoU ↑, H ↓, V ↓).
• Accurate surface roughness recovery.
• Strong zero-shot generalization.

Reviewer’s Takeaway

This work demonstrates how conditional latent diffusion models enable scalable, high-fidelity, real-time optical monitoring in LPBF. By integrating probabilistic generative modeling, latent compression, wavelet-based texture evaluation, and 3D geometric reconstruction, the framework advances beyond conventional super-resolution toward physics-aware additive manufacturing monitoring. The combination of distribution modeling and geometric validation makes this study a strong reference for researchers working on:

• In-situ LPBF monitoring
• Generative models in manufacturing
• Diffusion-based super-resolution
• Closed-loop defect detection systems

Slide deck: Download PDF

Core Contributions and Methodology

The paper introduces the Dynamic Segmentation Convolutional Neural Network (DSCNN), a pixel-wise semantic segmentation model designed specifically for high-resolution powder-bed imaging data. Unlike earlier patch-based approaches (e.g., MsCNN), DSCNN performs classification at the native image resolution and explicitly captures multi-scale contextual information through a three-leg architecture: (i) a global CNN branch for large-scale bed-level conditions, (ii) a regional U-Net branch for medium-scale morphological features, and (iii) a localization branch operating on native-resolution tiles to preserve fine-grained details. In addition, normalized pixel-coordinate channels are incorporated to encode spatial priors, reflecting the fact that certain defects are more likely to occur in specific regions of the build plate.

Data, Training Strategy, and Practical Considerations

DSCNN is validated across six different powder-bed AM machines spanning three technologies (laser PBF, electron-beam PBF, and binder jetting), demonstrating strong cross-machine generalization. The authors carefully address key challenges in AM data, including extreme class imbalance and noisy manual annotations. To mitigate these issues, the training pipeline employs median-frequency class balancing and a skeptical (hard-bootstrapping) loss, which allows the model to partially rely on its own confident predictions when human labels are inconsistent. A tile-based training strategy is used to manage GPU memory while preserving multi-scale context, enabling efficient training on very large images.

Results and Insights

Experimental results show that DSCNN significantly outperforms previous patch-based CNN approaches in terms of spatial localization accuracy and false-positive rates, particularly for rare but critical defect classes such as incomplete spreading, recoater streaking, and debris. Transfer learning experiments further demonstrate that models trained on data-rich machines can be effectively adapted to machines with limited labeled data, improving convergence speed and overall performance. Qualitative visualizations confirm that DSCNN produces more precise and interpretable anomaly maps, supporting its suitability for real-time, in-situ monitoring.

Reviewer’s Takeaway

This paper represents a strong example of domain-aware deep learning, where network architecture, input representation, and loss design are tightly aligned with the physical and operational characteristics of powder-bed AM processes. Its emphasis on pixel-wise segmentation, multi-scale context, and transferability across machines makes DSCNN a foundational reference for researchers working on real-time quality assurance and closed-loop control in additive manufacturing.

Slide deck: Download PDF