Engineers and scientists usually have to select a separate computational tool for each physical process: one simulator for water flow, another for how a protein folds. Each such program is manually configured for specific equations. A group of researchers from the Massachusetts Institute of Technology (MIT), the University of British Columbia (UBC) and other organizations proposed replacing this patchwork with a single system based on deep learning - WorldParticle.
The approach is based on representing any material as a set of tiny three-dimensional points, or particles. This way of describing things is called a Lagrangian representation: the system tracks the position and velocity of each particle in time and space. The neural network receives as input the coordinates and physical properties of the particles and predicts their state at the next step - roughly one second of model time later.
"The architecture is similar to ChatGPT, but instead of predicting the next word, we predict how the particles will move by one step," explains Peter Yichen Chen, a former MIT postdoc and now an assistant professor at UBC and one of the corresponding authors of the work.
Before each step, the model separately accounts for external forces, for example the impulse with which a wave strikes a wall. After prediction, the system returns these forces to the calculation and corrects the result. Thanks to this, WorldParticle can work with materials, conditions and loads that were not present during training.
To handle millions of particles, the researchers applied a method of merging and compressing particles while preserving fine details of behavior. On six types of physical phenomena - from deformable objects to granular materials and proteins - the model showed higher accuracy than numerical methods. In particular, it correctly described how a mass of sand settles into a pile, whereas some baseline methods left particles "hanging" in the air. And WorldParticle modeled the change in protein structure over time roughly 100 times faster than numerical simulators.
The authors emphasize: this is still a first step, not a finished product. In the future, they hope that a single model will help study new materials and phenomena that are difficult to investigate with existing methods. The work is published as a preprint on the arXiv server and will be presented at the SIGGRAPH Asia 2026 conference in Kuala Lumpur (Malaysia) on December 1-4.