Session: 03-01-01: Advanced Manufacturing

Paper Number: 188413

188413 - Uncertainty-Aware and Interpretable Transfer Learning for Digital Certification of Fatigue Properties in Additively Manufactured Alloys 

Additively manufactured (AM) metallic components allow for the creation of complex geometries but often show significant variability in their fatigue performance. This variability can be attributed to factors such as surface roughness, residual stresses, and heterogeneous defect morphologies. In safety-critical applications, it is essential to achieve reliable fatigue predictions that include calibrated uncertainty, physics-consistent behavior, and interpretable outputs that align with mechanical understanding and certification requirements. Conventional data-driven approaches often fail to explicitly address these key requirements, thereby limiting their trustworthiness and deployment in certification contexts. To overcome these limitations, we propose a multi-modal approach that directly learns from raw micro-CT data and loading parameters. This method incorporates physics regularization to enhance interpretability and reports epistemic uncertainty using Monte Carlo Dropout. Our approach prioritizes making interpretable, physically consistent predictions with quantified uncertainty, which is critical for individualized fatigue life prediction. This alignment with emerging digital certification workflows further enhances its applicability.

Our framework employs a unified, physics-guided, multi-modal neural architecture that couples micro-CT imaging with loading parameters for end-to-end fatigue life prediction. At its core, the model integrates an attention-guided, multi-task U-Net that simultaneously performs specimen segmentation and fatigue life regression from CT data, together with a compact loading-parameter branch that encodes physically meaningful trends in the S–N curve response. Building on the PPgNN 2.0 baseline, which has previously demonstrated superior performance over conventional neural network models for S–N curve modeling. The proposed architecture extends this capability by directly leveraging raw micro-CT imagery through spatial supervision and late-stage attention mechanisms.

To address data scarcity and reduce the reliance on extensive CT imaging, the loading-parameter branch is initialized via transfer learning from the PPgNN 2.0 model pre-trained on large, open-literature fatigue datasets. The transferred loading pathway is then fine-tuned jointly with the image branch using in-house CT data. This selective transfer strategy enables the model to exploit abundant non-imaging fatigue data while minimizing the number of costly, high-resolution CT scans required for training. Such an approach is critical for digital certification applications, where comprehensive fatigue testing with full CT characterization is prohibitively expensive to scale.

Architecturally, the image pathway is optimized using a composite objective that couples regression and spatial supervision to promote morphology-sensitive feature learning under limited data. Specifically, a regression loss on logarithmic fatigue life is combined with segmentation losses (e.g., Dice and cross-entropy) and an attention-consistency regularizer, encouraging the network to localize fatigue-relevant regions while respecting specimen boundaries. In parallel, physical consistency is enforced in the loading branch through weight and bias constraints that preserve monotonicity with respect to the stress ratio and curvature in the S–N curve relationship with respect to the stress amplitude. These constraints are implemented using sign-limited parameters, following established PPgNN formulations.

The proposed transfer learning procedure further enhances robustness under data sparsity by preserving physically meaningful trends learned from the literature while allowing the image pathway to adapt to specimen-specific morphology. Finally, epistemic uncertainty is quantified using Monte Carlo Dropout by performing multiple stochastic forward passes at inference and aggregating the resulting predictive distribution. This approach provides practical uncertainty intervals without introducing an explicit variance head, as commonly used in Gaussian mixture formulations.

Our unified, physics-regularized multi-modal model outperforms the PPgNN 2.0 baseline in predictive accuracy and generalization. The attention-guided U-Net highlights fatigue-critical regions in CT images, while the physics-constrained loading branch ensures the S-N curve predictions are physically meaningful. Transfer learning allows robust adaptation to new loading conditions, and MC Dropout provides calibrated uncertainty bands for risk-aware decision-making. However, limitations include a focus only on epistemic uncertainty, descriptive attention maps that don't establish causality, and physics constraints limited to monotonicity and curvature. Future efforts will aim to expand physics constraints, incorporate other uncertainties, and validate the approach with a wider variety of materials and loading conditions. Despite these challenges, our results represent significant progress toward interpretable, physics-consistent, and uncertainty-aware machine learning for the digital certification of additively manufactured components.
 

Presenting Author: Yongming Liu Arizona State University

Presenting Author Biography: Yongming Liu is a professor of aerospace and mechanical engineering with the School for Engineering of Matter, Transport and Energy at Arizona State University. He heads the Prognostic Analysis and Reliability Assessment Laboratory (PARA). He joined ASU in 2012.

His research interests range from fatigue and fracture of engineering materials and structures, probabilistic computational mechanics, risk assessment and management, to multi-physics damage modeling and structural durability, multi-scale uncertainty quantification and propagation, imaging-based experimental testing, diagnostics, and prognostics.