WorldParticle Brings One Simulation Design to Sand Fluids and Proteins

Researchers at the Massachusetts Institute of Technology (MIT), the University of British Columbia (UBC) and collaborating institutions have developed WorldParticle, a system for learning how materials move. In a November 9 account, MIT described applications ranging from collapsing sand to changing protein structures. The work offers a common simulation architecture for problems that ordinarily require substantial specialist engineering. Its preprint first appeared in May. 1,2

The researchers trained a separate network for each of six categories, including cloth, solids, fluids and molecular dynamics. Each uses the same architecture to accept particles and their properties and predict movement. A single set of weights spanning all six categories remains a future research goal. 2

Representing matter as particles gives the method a common language. Each point carries information such as position and velocity, while the model learns relationships among points and nearby boundaries. Peter Yichen Chen, a UBC researcher and co-corresponding author, described the process in MIT's account: “we are predicting how these particles will move, one step at a time.” Repeating that prediction produces a moving simulation. 1

For an engineer, the attraction is less rebuilding of software when the material changes. A team developing a soft gripper, for example, may need to understand how its shape responds to contact before choosing a design to manufacture. Reusing a simulation architecture could let the team investigate more candidates with a consistent set of tools. The usefulness of any prediction would still depend on how well it captures the material and contact conditions that the device will encounter.

WorldParticle first advances particles using known external forces. A neural network then adjusts the prediction for interactions among particles and with their surroundings. To manage computation, it compresses particle information into a smaller collection of summaries, exchanges information there, and returns corrections to individual particles. Local contact and broader motion can influence the same prediction. 2

Learning to simulate particle motion has a substantial research history. A 2020 paper by Alvaro Sanchez-Gonzalez and colleagues represented particles as points in a graph and learned their interactions through messages passed along connections. That work demonstrated transfer to different starting conditions and longer simulations. WorldParticle builds within this wider effort to make learned simulators useful beyond the particular examples used to train them. 3

The paper reports tests with changed materials, geometry and forces. Memory limits the current implementation to about 50,000 particles on one A100 graphics processor with 80 gigabytes of memory. Training used established numerical solvers, alongside separate experiments with reconstructed real-world motion. The method's reach depends on its training evidence and the conditions being predicted. 2

For scientific and engineering use, the next question is how the predictions support a decision. A visually convincing collapse of sand may be enough for an animation; selecting a load-bearing component requires evidence about the quantities that determine whether it will work. Comparisons with measurements, the behavior of errors over long simulations and the effort needed to train a new material would show where this shared architecture can save useful work.

Sources

1. MIT News account and Peter Yichen Chen quotation, November 9, 2026

2. Wang, Guo, Chen and colleagues, WorldParticle paper, especially Sections 3.4, 4 and 5; preprint first submitted May 14, 2026

3. Sanchez-Gonzalez and colleagues, Learning to Simulate Complex Physics with Graph Networks, 2020