Researchers at MIT, the University of British Columbia and other institutions reported on Nov. 9, 2026, a single deep-learning system called WorldParticle that represents materials as particles and predicts their movement step by step. Instead of coding a separate solver for each process, it learns underlying dynamics, which the team says lets it adapt to materials, forces and environments it did not encounter in training.

In tests, the researchers modeled six physical phenomena, including deformable objects, granular materials and proteins, reporting higher accuracy than numerical methods on those cases. They said WorldParticle modeled protein structural change about 100 times faster than numerical simulators. The work was described in an open-access paper and was scheduled for presentation at an ACM SIGGRAPH conference in Asia.

Co-corresponding author Peter Yichen Chen, a former MIT postdoc now at UBC, described the effort as a first step toward a general simulation foundation model. The team said future work would aim to adapt the system to soft-robot manipulation, RNA evolution and astrophysical phenomena such as galaxy mergers, and to build a broader simulation pipeline for scientific discovery.