UniPhy is a unified latent-conditioned neural model which learns a common latent space to encode the properties of diverse materials. At inference, given motion observations for a system with unknown material parameters, UniPhy allows material inference via differentiable simulation-based latent optimization.
We propose UniPhy, a common latent-conditioned neural constitutive model that can encode the physical properties of diverse materials. At inference UniPhy allows `inverse simulation' i.e. inferring material properties by optimizing the scene-specific latent to match the available observations via differentiable simulation. In contrast to existing methods that treat such inference as system identification, UniPhy does not rely on user-specified material information. Compared to prior neural constitutive modeling approaches which learn scene specific networks, the shared training across materials improves both, robustness and accuracy of the estimates. We train UniPhy using simulated trajectories across diverse geometries and materials -- elastic, plasticine, sand, and fluids (Newtonian & non-Newtonian). At inference, given an object with unknown material properties, UniPhy can infer the material properties via latent optimization to match the motion observations, and can then allow re-simulating the object under diverse scenarios. We compare UniPhy against prior inverse simulation methods, and show that the inference from UniPhy enables more accurate replay and re-simulation under novel conditions.
Material Point Method (MPM) is a hybrid Euler-Lagrangian method for particle-based simulation. MPM requires the material type and parameter information beforehand to compute the deformation gradient projection and stress via hand-designed analytical model. Additionally, each material has its own analytical model. We propose a unified, neural model UniPhy, that can learn representations across diverse materials without the need for human-in-the-loop to define material information beforehand. We achieve this by learning neural representations of materials through a latent space.
Given a trajectory observation, we optimize the latent which encodes the material properties of the object in the scene. Using this optimized latent, we can resimulate it under new configurations. Click on the different material types to check out more examples.
We interpolate the latent codes among different material types. Use the sliders below to interpolate between the material.
@article{mittal2025uniphy,
title={UniPhy: Learning a Unified Constitutive Model for Inverse Physics Simulation},
author={Mittal, Himangi and Zhuang, Peiye and Lee, Hsin-Ying and Tulsiani, Shubham},
journal={arXiv preprint arXiv:2505.16971},
year={2025}
}