MIT develops unified architecture to simulate fluids, grains, and proteins

Researchers at MIT have developed a unified architecture capable of simulating fluids, granular materials, and proteins within a single system.

Conventional computational physics relies on dedicated numerical solvers tailored to specific states of matter. Navier-Stokes formulations govern continuous fluid flows, discrete element methods calculate collisions between rigid grains, and molecular dynamics engines resolve atomic forces in macromolecular structures. These bespoke methods require distinct mathematical approximations, carry heavy computational overheads, and scale poorly when systems transition across phases.

The MIT team bypassed specialised partial differential equation solvers by adopting a unified Lagrangian perspective. By treating matter as interacting particle systems across scales, the architecture models state trajectories using a shared graph-based formulation. Nodes represent individual particles or coarse-grained clusters, while dynamic edges encode pairwise forces and spatial relationships.

The system predicts velocity updates and position shifts over discrete time steps. Because the model learns underlying interaction principles rather than relying on hardcoded constitutive laws, it generalises across disparate physical regimes. During benchmarking, the framework reproduced the splashing of liquids, the shear jamming of sand, and the conformational changes of peptide chains without requiring structural retuning between tasks.

In comparative evaluations, the architecture generated rollouts orders of magnitude faster than traditional numerical integration schemes. The model maintained stable trajectories over multi-step rollouts, preserving bulk macroscopic properties such as mass and volume conservation. By learning latent dynamics directly from particle trajectories, the system demonstrates that distinct macroscopic and microscopic behaviours share common computational representations.

Key constraints remain around long-horizon autoregressive error accumulation and exact energy conservation. The researchers plan to scale the model to larger particle counts and introduce inductive biases that enforce strict thermodynamic consistency. Can a single learned representation ever fully replace the formal analytical proofs underpinning classical numerical physics?