Attention-guided training: Integrating Domain Knowledge into Explainable Deep Learning for Crack Tip Segmentation

Kurzfassung

Deep Learning (DL) has significantly advanced numerous scientific disciplines in recent years, driven by its ability to learn complex, non-linear patterns from data. However, DL models often lack transparency and are commonly perceived as black-box systems. Without appropriate explainability, their predictions remain difficult to interpret and trust. This significantly hinders their application—particularly in high-stakes applications such as fatigue crack detection in aerospace components. Recent studies have demonstrated the potential of AI-based semantic segmentation for detecting cracks in digital image correlation (DIC) data acquired during mechanical testing. Additionally, post-hoc eXplainable AI (XAI) methods such as Grad-CAM have been successfully employed to interpret U-Net-based segmentation results in this context, not only demystifying the model’s decision-making process but also improving generalization by enabling the selection of models whose attention patterns align with domain knowledge. Complementary approaches aim to embed domain knowledge directly into the learning process through physics-guided machine learning techniques, where models are constrained by known physical laws or boundary conditions expressed as equations. Although recent studies use explanations to enhance the training process (e.g., by using a secondary feedback DL model) domain knowledge is often applied indirectly - e.g., by experts evaluating whether model attention patterns are physically plausible—a framework that enables this knowledge to be incorporated explicitly, in the form of physical equations guiding model attention during training, is still lacking. In this study, we introduce Attention-Guided Training (AGT) — a framework that integrates XAI with domain knowledge to address both the faithfulness of XAI and model robustness. AGT enables domain knowledge to directly guide model attention during training. We demonstrate the approach on the task of AI-based crack tip segmentation in DIC data, where domain knowledge is readily available in the form of mathematical equations describing the physical displacement field induced by a single crack. Our results show that AGT improves both generalization and trustworthiness.