MIT Team Develops New Simulator to Predict Particle Motion in Six Physical Systems

Researchers from the Massachusetts Institute of Technology (MIT) and the University of British Columbia (UBC) have developed a transformer-based simulator design capable of predicting particle motion in six types of physical systems, including sand and proteins, according to a paper and a release from MIT. Traditional physical simulators are typically built for specific tasks, with the equations used by fluid solvers differing from those used by protein folding solvers, and each solver requires manual coding. The new system, named WorldParticle, aims to replace these independent programs with a learned design, although each type of material will still have its own training network.

WorldParticle treats each material as a group of particles, tracking each particle based on its position and velocity, and gradually predicting its next position. UBC Assistant Professor and co-corresponding author Peter Yichen Chen likened this approach to ChatGPT, which is a system that predicts text word by word.

How the WorldParticle Simulator Works

The team divided each step into two parts in a preprint on arXiv. First, basic calculations are performed to move particles based on known external forces (such as driving forces). Then, positions and velocities are adjusted through a learned corrector to account for interactions between particles. To reduce computational costs, the model merges similar particles into smaller summary tokens, which then relay information back to each particle.

Researchers trained independent networks for six categories (fabric, elastic solids, two fluids, sand, and protein motion), with training data sourced from existing solvers. According to the paper, the model exhibited lower errors and higher stability in long-duration runs compared to three other neural simulators in fluid, fabric, and sand tasks. These competitors produced floating sand particles and exploding fabric during long simulations. The time step for protein motion is 50 femtoseconds, which is 100 times the 0.5 femtosecond step used in traditional molecular simulations. A femtosecond is 10 to the power of negative 15 seconds.

Future Directions for the Model

Since the first step directly addresses external forces, the model can respond to thrusts not seen during training. The team has built a real-time tool that allows users to apply forces to soft objects. In three segments of real-world operation videos, the predicted motion aligns in mass with the observed motion, although the paper does not report error data. Several limitations also apply. The paper notes that memory is a primary constraint: the first layer views each particle simultaneously, limiting the scene to about 50,000 particles and requiring a single 80GB graphics processor.

The authors also stated that future work will focus on sharing a set of weights across all categories, and that the model's accuracy may be compromised compared to dedicated solvers. "We hope to achieve a GPT moment in simulation," Professor Chen stated, describing a foundational model that could apply to any scenario. The team plans to adapt WorldParticle to more physical systems, including flexible robot tasks, and move toward a general simulation pipeline for scientific discovery, MIT stated.

The Impact of WorldParticle on Physical Simulation

The development of the WorldParticle simulator marks a significant advancement in physical simulation technology, enabling effective predictions of particle motion across various physical systems through a transformer-based design. This not only enhances the accuracy of simulations but also reduces computational costs, particularly in terms of stability during long-duration runs. As future plans aim to apply this technology to more physical systems, it may play a crucial role in flexible robotics and other scientific research fields, driving further advancements in related technologies.