Risk-Aware Decision Making for Metal Additive Manufacturing via Conditional Diffusion Models

Laser Powder Bed Fusion (LPBF) is a leading technology for producing complex metal components with high performance, but process outcomes are highly sensitive to manufacturing conditions. Key challenges include a high-dimensional process parameter space (e.g., laser power, scan speed, chamber temperature, hatch spacing), stochastic defect formation (e.g., recoater-related defects, incomplete spreading, swelling, debris, porosity), and time- and cost-intensive experimental validation. Existing works mainly rely on generative modeling and defect prediction in isolation, and no generative modeling has previously been applied on the Peregrine dataset.

To address these limitations, we propose a unified Generator–Evaluator–Decision framework that integrates generative modeling and defect prediction into a real-time decision-support system capable of (1) replacing costly physical experimentation with probabilistic print forecasting, (2) estimating defect distributions, and (3) supporting risk-aware manufacturing decisions in situ.

Framework Overview

The proposed framework consists of three components:

CDM-based Generator. Given a context window of \(k\) observed layers up to layer \(i\), a Conditional Diffusion Model (CDM) generates the next-layer post-melt image patch by learning \(p_\theta(X_{i+1}[x_0,y_0] \mid \text{context}_{i-k+1:i}(x_0,y_0))\). The conditioning context fuses past post-melt and post-recoat image history (encoded via CNN+GAP), temporal layer-wise signals (tokenized by a lightweight Transformer), the target-layer scan heatmap, and process-parameter masks. Two complementary conditioning pathways inject context tokens into the U-Net bottleneck (for global structure) and spatial maps as additional inputs (for local process information). The model generates multiple plausible realizations \(X_0^{(1)}, X_0^{(2)}, \ldots, X_0^{(M)}\) through iterative denoising from Gaussian noise.

CNN-based Defect Evaluator. The evaluator maps each generated sample to defect-related outcomes: \(Y^{(m)} = f_\theta(X^{(m)})\). Taking a 2-channel tensor of [generated post-melt, post-recoat] as input, it outputs defect rates for 12 defect classes per patch. The set of predictions \(\{Y^{(1)}, Y^{(2)}, \ldots, Y^{(M)}\}\) is aggregated to form an empirical defect distribution, enabling distribution-level risk estimation rather than single deterministic predictions.

Rule-based Decision Engine. The estimated defect distribution is used to support real-time manufacturing decisions via a user-defined risk threshold \(D^{\text{threshold}}\):

• “Keep printing” if predicted defect risk \(\leq D^{\text{threshold}}\)
• “Stop printing” if predicted defect risk \(> D^{\text{threshold}}\) (early termination to avoid further production cost)

Dataset

This work uses the Peregrine digital twin dataset released by Oak Ridge National Laboratory (DOI: 10.13139/ORNLNCCS/2008021), which consists of 5 LPBF builds with all modalities spatially (pixel-wise) and temporally (layer-wise) aligned: post-melt images, post-recoat images, part ID maps, anomaly segmentation masks (12 classes), laser scan paths, temporal in-build signals, and part-level process parameters. The primary training set uses Builds 1, 2, and 5 (aligned geometry and consistent physical meaning), with Builds 3 and 4 reserved for out-of-distribution evaluation.

Raw HDF5 scan paths, part IDs, and process parameters are converted into pixel-aligned masks to create a unified layer-aligned multimodal tensor dataset, enabling efficient training without repeated HDF5 traversal.

Results

The CDM generates high-quality virtual printings whose structure and texture closely match ground truth. Intensity distributions of the generated samples are well-aligned with ground truth distributions. The defect distribution estimated from the generated samples captures the ground-truth defect rates in the mean-period across 12 defect classes. Real-time GO/NO-GO decisions can be triggered layer-by-layer: for example, if the 100th real printing yields a 2nd-class defect rate of 0.230, printing is halted as this defect will propagate and worsen through subsequent layers.

Contributions

• A first unified “Generator–Evaluator–Decision” framework for risk-aware LPBF monitoring
• A first multimodal Conditional Diffusion Model applied to the Peregrine digital twin dataset, fusing image history, temporal signals, scan paths, and process parameters
• Distribution-aware decision-making and uncertainty quantification via defect distributions rather than point predictions, supporting real-time GO/NO-GO decisions under uncertainty

Conference: Download Slides (PDF)\

Poster: Download Poster (PDF)

Video: Download Poster (MP4)

abstract: Download Poster (PDF)

Citation:
Ko, Donghyun; Peloquin, Jacob (2026). Risk-Aware Decision Making for Metal Additive Manufacturing via Conditional Diffusion Models. Presented at IISE Annual Conference & Expo 2026.