Graph tool tracks changes in physics learning inside scientific models
A Graph-Based Inspection and Intervention Tool for Assessing Mechanistic Learning in PINNs
Machine LearningArtificial Intelligence
Summary
Sometimes, scientific AI models learn to represent physical concepts during training, but it's unclear if these learned connections stay meaningful when the model faces new situations. The authors created GIIT, a tool that maps parts of a model to elements of known physics and tests whether this mapping holds when the input data changes over time. They found that as models make predictions beyond their training range, the internal connections to physics shift and become less stable, even before the model's overall accuracy falls. This can help identify when a model's understanding is drifting, potentially improving trust in AI-based scientific tools.
What this means in practice
- •For scientific machine learning engineers: Inspect and verify how physics-aware AI models internally represent physical laws when predicting outside their training data range.
- •For software developers for physics simulations: Use GIIT to monitor and debug physics-informed models to maintain meaningful physical correspondences in extended simulation scenarios.
Authors
Adwait Patkhedkar, Alifaraz Lakhani, Prathmesh Mohite, Abhijeet Salunke
Abstract
We ask whether physically meaningful correspondences discovered inside a trained scientific model remain meaningful outside the conditions under which they were discovered. We introduce GIIT (Graph-based Inspection and Intervention Tool), which represents governing physics as a computational physics dependency graph, maps graph nodes to internal network components via sensitivity- and trend-based discovery, and tests the resulting mapping under targeted intervention. Evaluating on temporal extrapolation out-of-distribution (OOD) regimes of increasing severity, we find that temporal extrapolation is associated with layer-wise correspondence drift and reduced functional correspondence. Specifically, while the model achieves low physics residual in-distribution (1.915 x 10-4 on full ID and 1.01 x 10-4 on an ID sub-window t in [0.5, 0.8]), the functional mapping changes even before leaving the training domain, and the shift continues as the evaluation window extends beyond the training domain: residual error increases from 2.81 x 10-2 on the boundary-crossing window (t in [0.5, 1.5]) to 3.30 x 10-1 on severe OOD (t in [1, 2]), accompanied by a systematic leftward shift of internal layer mappings, where the average winner layer index drops from 5.71 (ID) to 3.00 (ID sub-window) down to 2.29 (severe OOD). Only 2 of 7 physical nodes maintain stable layer assignments across the boundary-crossing window, and only 1 of 7 under severe extrapolation, revealing potential internal functional correspondence changes before severe degradation in conventional output metrics. Additional results for linear oscillatory systems are further detailed in the appendix.