Bridging model attention to physical insights in structure-property dynamics

Time-resolved experiments increasingly capture structural evolution together with measured physical responses, creating new opportunities to interrogate spatiotemporal structure–property dynamics. Although deep learning models, including attention-based transformers, can predict such responses directly from image sequences by modeling complex correlations, predictive accuracy alone does not provide physical understanding. Attention mechanisms identify statistically influential regions within evolving images, yet these signals do not explicitly reveal the structural organization governing system response. A framework is therefore established in which attention–derived emphasis guides physics-informed structural analysis, treating attention not as explanation but as a constraint for domain-grounded transformation that exposes the dominant mode of structural organization associated with measured properties. Applied to gold-nanocontact deformation (atomic scale) and wet rubber–surface contact (macroscale), the framework recovers, in each system, the structural organization associated with the measured response. By translating attention–derived emphasis into physical descriptors, this work establishes a pathway from predictive modeling to physical interpretation of hidden structure–property relationships.