Trustworthy AI-based crack-tip segmentation using domain-guided explanations

TL;DR

This work tackles the trustworthiness of deep learning in crack-tip semantic segmentation from digital image correlation data by introducing Attention-Guided Training (AGT), which couples a prediction objective with a domain-informed explanation objective. Explanations are generated via CAM-based methods and steered toward a Williams-based near-tip stress-field prior, operationalized through target maps derived from the von Mises stress field $\sigma_{\rm VM}$ and Williams coefficients. The authors compare CAM variants under the Co12 criteria and show gradient-based CAMs provide the best faithfulness for segmentation. They demonstrate that AGT with a physically grounded target (Binary Williams or Gradual Williams) improves Dice-based performance, reliability on out-of-distribution data, and the faithfulness of explanations, with Binary Williams delivering the strongest overall gains. Overall, the paper provides a principled framework to fuse explainability, objective evaluation, and domain knowledge to enhance both predictive reliability and interpretability in a high-stakes scientific task, with potential generalization to other domains such as materials science and structural integrity assessment.

Abstract

Ensuring the trustworthiness and robustness of deep learning models remains a fundamental challenge, particularly in high-stakes scientific applications. In this study, we present a framework called attention-guided training that combines explainable artificial intelligence techniques with quantitative evaluation and domain-specific priors to guide model attention. We demonstrate that domain-specific feedback on model explanations during training can enhance the model's generalization capabilities. We validate our approach on the task of semantic crack tip segmentation in digital image correlation data, which is a key application in the fracture mechanical characterization of materials. By aligning model attention with physically meaningful stress fields, such as those described by Williams' analytical solution, attention-guided training ensures that the model focuses on physically relevant regions. This finally leads to improved generalization and more faithful explanations.

Figures (9)

• Figure 1: Overview of attention-guided training (AGT); AGT builds upon conventional machine learning principles, providing enhancements through explainable artificial intelligence (XAI) techniques. It incorporates post-hoc explanations and objective XAI evaluation metrics to assess model explainability and integrates domain knowledge through a modified training loop, ensuring better alignment between trained patterns and expert insights.
• Figure 2: Semantic differences in model explanations, calculated using Grad-CAM, despite similar predictions. a) The explanation of the U-Net model prediction shows that the model identifies the crack path to predict the crack tip location. It is known that experimental data in this region likely contains artifacts from the digital image correlation evaluation (see Section \ref{['sec:Data']} for the experimental context). b) In contrast, the explanation of the ParallelNets MelchExplainCrack model prediction highlights the region ahead of the crack tip as most relevant for its prediction This attention aligns closely with domain knowledge shown in c), where the crack tip stress field is accurately represented by the widely accepted Williams series (Equation \ref{['eq:williams_stress']}), thus relating model behavior directly to theoretical foundations in fracture mechanics.
• Figure 3: Illustration of the exemplary DL task. We obtain data through full-field DIC analysis during fatigue crack growth experiments (left). The resulting displacement fields are input into a segmentation CNN (here, U-Net architecture), which is trained to segment the most likely crack tip pixels. The centroid of the predicted segmentation yields the crack tip position, which is then translated back into the reference frame of the sample surface.
• Figure 4: Visualization of explanations and performance comparison: The top row displays attention heatmaps generated using various CAM techniques, each corresponding to the same prediction made by ParallelNets on a sample of the $S_{160,4.7,val}$ dataset (see Section \ref{['sec:Data']}). These heatmaps are based on the encoder layers of the U-Net architecture. Given the significant visual differences between methods, the accompanying bar plot quantifies their relative performance based on the four criteria --- correctness, completeness, continuity, and compactness.
• Figure 5: Example of an AGT for crack tip segmentation. The Dice loss provides feedback on the segmentation quality, while the cosine similarity loss quantifies coherence with the target explanation (e). Plots (a) and (b) show the loss curves of the segmentation quality and explanation coherence, respectively. Plot (c) shows the attention heatmap after the pretraining phase of 30 epochs. Plot (d) depicts the heatmap of the model with the smallest validation loss during the AGT phase (indicated by a dashed orange line in (a, b)). All attention heatmaps were obtained using the Grad-CAM++ method on the encoder layers of the U-Net.