This paper was accepted on the AI for Science Workshop at NeurIPS 2025.
Simulating interactions between deformable our bodies is significant in fields like materials science, mechanical design, and robotics. Whereas learning-based strategies with Graph Neural Networks (GNNs) are efficient at fixing advanced bodily techniques, they encounter scalability points when modeling deformable physique interactions. To mannequin interactions between objects, pairwise world edges need to be created dynamically, which is computationally intensive and impractical for large-scale meshes. To beat these challenges, drawing on insights from geometric representations, we suggest an Adaptive Spatial Tokenization (AST) technique for environment friendly illustration of bodily states. By dividing the simulation house right into a grid of cells and mapping unstructured meshes onto this structured grid, our method naturally teams adjoining mesh nodes. We then apply a cross-attention module to map the sparse cells right into a compact, fixed-length embedding, serving as tokens for the whole bodily state. Self-attention modules are employed to foretell the subsequent state over these tokens in latent house. This framework leverages the effectivity of tokenization and the expressive energy of consideration mechanisms to attain correct and scalable simulation outcomes. In depth experiments show that our technique considerably outperforms state-of-the-art approaches in modeling deformable physique interactions. Notably, it stays efficient on large-scale simulations with meshes exceeding 100,000 nodes, the place current strategies are hindered by computational limitations. Moreover, we contribute a novel large-scale dataset encompassing a variety of deformable physique interactions to help future analysis on this space.
