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

Domain expertise improves AI reliability: The study shows that incorporating domain-specific knowledge (in this case, physics-based stress field models) during AI model training leads to better generalization and more trustworthy predictions, suggesting that AI systems perform better in high-stakes applications when they align with established scientific principles.

Explainability and performance go hand-in-hand: Rather than treating model explanations as an afterthought, the researchers used explainable AI techniques actively during training to guide what the model "pays attention to," demonstrating that more interpretable AI can also be more accurate and robust.

AI explanations should be physically meaningful: The framework ensures that when the AI model makes decisions, it focuses on regions that have actual physical significance (stress concentrations around crack tips) rather than spurious patterns, making the model's reasoning scientifically defensible and easier for experts to validate.

Quantitative evaluation of trustworthiness is possible: The paper demonstrates a concrete methodology for measuring whether an AI system is trustworthy in scientific applications—not just through accuracy metrics, but by verifying that the model's internal reasoning aligns with known physical laws and expert understanding.

Responsible AI requires interdisciplinary collaboration: Developing trustworthy AI for specialized fields like materials science requires combining machine learning expertise with domain knowledge, suggesting that responsible AI development cannot be purely technical but must involve subject matter experts throughout the process.