Using explainability to design physics-aware CNNs for solving subsurface inverse problems

Explainability tools can improve AI design beyond accuracy alone: By using explainability techniques (Score-CAM and Deep SHAP), researchers designed a physics-aware CNN that achieved comparable predictive accuracy to deeper networks while improving "descriptive accuracy"—meaning the model's internal logic was more interpretable and aligned with physics principles.

Simpler models can outperform complex ones when properly designed: The study found that a shallow CNN with just two convolutional layers outperformed a deeper Encoder-Decoder network for subsurface imaging, challenging the common assumption that "bigger is better" and demonstrating that explainability-guided design can lead to more efficient solutions.

XAI methods can guide hyperparameter selection toward trustworthiness: Rather than relying solely on automatic optimization algorithms that prioritize prediction accuracy, the researchers used explainability methods to select hyperparameters (like kernel sizes and network depth), resulting in a model whose decision-making process was more transparent and aligned with known physics.

Domain expertise and AI transparency should work together: This approach bridges the gap between machine learning and domain knowledge by using explainability to ensure the CNN "thinks" in ways consistent with subsurface physics, rather than discovering arbitrary patterns that happen to work but lack scientific grounding.