<p>Predicting the full-range stress–strain behavior of architected elastomers is challenging due to limited data and complex deformation mechanisms. We propose a physics-informed multimodal machine learning framework that embeds domain knowledge at both the data and model levels. At the data level, the Gibson–Ashby law guides targeted augmentation, and energy absorption efficiency segments the curves into distinct stages. At the model level, a multi-branch attention architecture dynamically fuses multimodal inputs and quantifies stage-specific contributions. Our framework achieves high accuracy (R² = 0.921), and the learned attention weights reveal a transition in mechanism from structure-dominated collapse to material-dominated densification. This work establishes an interpretable platform for uncovering structure–property relationships and deformation mechanisms in mechanical metamaterials.</p>