Neural Modular Physics for Elastic Simulation

Learning-based methods have made significant progress in physics simulation, typically approximating dynamics with a monolithic end-to-end optimized neural network. Although these models offer an effective way to simulation, they may lose essential features compared to traditional numerical simulators, such as physical interpretability and reliability. Drawing inspiration from classical simulators that operate in a modular fashion, this paper presents Neural Modular Physics (NMP) for elastic simulation, which combines the approximation capacity of neural networks with the physical reliability of traditional simulators. Beyond the previous monolithic learning paradigm, NMP enables direct supervision of intermediate quantities and physical constraints by decomposing elastic dynamics into physically meaningful neural modules connected through intermediate physical quantities. With a specialized architecture and training strategy, our method transforms the numerical computation flow into a modular neural simulator, achieving improved physical consistency and generalizability. Experimentally, NMP demonstrates superior generalization to unseen initial conditions and resolutions, stable long-horizon simulation, better preservation of physical properties compared to other neural simulators, and greater feasibility in scenarios with unknown underlying dynamics than traditional simulators.

Computer simulations help scientists and engineers predict how physical objects move, bend, and deform. Recent AI-based simulators can be fast and flexible, but they often work like a single black box: they predict the final motion without revealing the physical steps in between. This makes them harder to understand, harder to control, and less reliable over long simulations. This paper introduces Neural Modular Physics, an AI simulator designed to work more like a traditional physics simulator. Instead of learning the whole simulation in one piece, it breaks the process into meaningful parts, such as estimating internal material forces and updating the object’s motion. This structure makes the model easier to inspect, allows physical rules to be added directly, and lets learned components be swapped with traditional physics components when useful. Experiments on deformable objects show that this modular design produces more accurate and stable long-term simulations than several existing AI simulators. It also generalizes better to new starting conditions and higher-resolution objects, while remaining faster than traditional physics simulation.