← All articles

🤖 AI

One AI Simulates Sand, Cloth, and Proteins. The Protein Runs 100x Faster.

MIT's WorldParticle replaces five separate physics solvers with one transformer architecture, and our math on its published numbers shows protein motion running 14.7 times faster than real time on a single GPU.

Illustration: a single glowing neural network above a split scene, sand dunes collapsing into particles on one side and a protein molecule rendered as colored spheres on the other, connected by streams of light.

One hundred times faster. That is how much quicker MIT's new WorldParticle system models a changing protein structure than the numerical simulators researchers use today, according to MIT News. Our own math on the team's published numbers makes it stranger: protein motion simulated 14.7 times faster than real time, on a single GPU, with 100 simulated seconds of molecular movement finishing in 6.8 seconds while the classical baseline would have needed about 11 minutes.

WorldParticle comes from MIT, UBC, and collaborators at seven other institutions, announced November 9, 2026 ahead of a SIGGRAPH Asia presentation in December. Its trick is refusal to specialize. Today a scientist simulating a physical process must pick the right tool for that one job, a solver built for flowing water or another built to model how a protein folds, each hand-coded around the equations that govern that process. WorldParticle throws that patchwork out and replaces it with one transformer architecture that treats every material as particles and learns how those particles move.

Everything Is Particles

Whether the system is a dune of sand, a sheet of cloth, a splash of water, or a protein molecule, WorldParticle breaks it into thousands or millions of tiny three-dimensional dots and predicts where each dot goes next, one step at a time. It is, in the words of co-corresponding author Peter Yichen Chen, "a similar architecture to ChatGPT. But instead of predicting one word at a time to write a piece of text, we are predicting how these particles will move, one step at a time."

Each step runs in two parts. An explicit predictor first advances every particle under known external forces like gravity, capturing externally driven motion but not how particles push on each other. A learned corrector then predicts the residual position and velocity updates: a tokenizer encodes local interactions, a super-token encoder merges those tokens into a compact set by halving the count at each level, and a decoder lifts the result back to full resolution. Because the architecture never mentions a specific equation, each phenomenon becomes a different training distribution over the same input-output format, not a new model design, which is why the same code can in principle simulate a protein, a sand dune, and a sheet of cloth without anyone rewriting the solver for each one.

Our Math: Faster Than Real Time

MIT News reports that WorldParticle predicts each frame of motion in one-second increments, which means the paper's throughput table reads directly as a faster-than-real-time factor. Nobody in the paper or the press coverage ran this conversion, so we did. Here is what the team's own Table 1, measured on a single NVIDIA A100 with 80 GB of memory, actually implies.

WorldParticle throughput on a single NVIDIA A100 (80 GB), from Table 1 of arXiv:2605.15305.
MaterialFrames / secondFaster than real timeDegrees of freedomCompute per step
Cloth24.0524.1x29,4000.923 TFLOP
Fluid16.1816.2x42,1201.323 TFLOP
Protein14.7014.7x61,8840.944 TFLOP
Elastic solid11.7111.7x52,5421.651 TFLOP
Granular sand11.5011.5x60,0001.886 TFLOP
Non-Newtonian fluid11.1711.2x60,0001.886 TFLOP

Every category runs faster than real time. Cloth leads at 24 times real time. Protein sits at 14.7 times. Even the slowest, non-Newtonian fluid, runs 11.2 times faster than the physics it simulates. In wall-clock terms, a 600-frame fluid rollout, ten simulated minutes of water sloshing, finishes in 37 seconds. And the protein case is where the 100x claim lands: 100 simulated seconds in 6.8 seconds on WorldParticle, versus roughly 680 seconds, or 11.3 minutes, for the numerical baseline at one-hundredth the speed.

Five Solvers Become One Architecture

Count what WorldParticle replaces. Section 4 of the paper names the classical solvers that generated its training data: DFSPH for fluids, GIPC for cloth, MPM for sand and non-Newtonian fluids, VBD for elastic solids, and OpenMM for protein dynamics. Five codebases, each with its own discretization and equations, across six categories of physics. One fixed architecture now covers all six, and that consolidation is the real story, more than any single speedup. A coupled problem like oil sloshing against a deformable valve once needed three specialists. Now it can in principle run on one system.

Super-token merging is what makes the consolidation affordable. Full self-attention over the sand scene's 60,000 degrees of freedom would score 3.6 billion particle pairs in a single layer, a cost that grows with the square of the scene size. WorldParticle never pays that bill: it compresses particles into super tokens, halving the count at each level, and lets the decoder communicate through that compact set, so global information propagates in every correction step instead of crawling through dozens of local hops.

It Generalizes, and That Is the Surprise

Speed is cheap if a model only memorizes its training scenes. WorldParticle's more interesting claim is that it handles scenes it never saw. A tiger-shaped pile of sand, a shape nowhere in the training set, collapses stably with plausible granular flow. It worked. An 800-frame fluid rollout on an unseen container tracks the ground truth, splash patterns and long-term settling included, without drifting. Cloth trained on 200-frame sequences stays coherent at 250 frames, accumulating twist without diverging, which is exactly where autoregressive neural simulators usually fall apart.

Beyond forward simulation, the team demonstrates three downstream uses: interactive control, where a user applies forces of any direction, location, and magnitude to a deformable object in real time; inverse design, which recovers a friction coefficient by backpropagating through a rollout until a duck-shaped inflatable mesh slides down a ramp and stops at a prescribed target; and learning directly from real-world point-cloud observations, predicting manipulation trajectories without a classical simulator in the loop.

Limitations

This analysis rests on a preprint and a university press release, not a peer-reviewed paper or an independent benchmark, because WorldParticle is an arXiv preprint from May 2026, presented at SIGGRAPH Asia in December 2026, and no third party has replicated the 100x protein figure as of November 22. That figure comes from an author's remark to MIT News, not a timed head-to-head comparison in the paper. Treat it as a claim from the team, not a measured benchmark. MIT News says the system achieves higher accuracy than numerical methods, while the paper itself admits the single architecture comes "at the cost of some per-phenomenon accuracy compared to those specialized methods," and the paper's framing is the safer one. Memory caps the ambition at roughly 50,000 particles on a single 80 GB A100, far below the millions in a production scene like a full car crash, a whole vehicle in a crash test, or a solvated protein complex where water molecules outnumber the protein a hundred to one. Quality is uneven, too: protein position error is 516,585 times smaller than fluid position error in the paper's own Table 1. And WorldParticle learns from trajectories generated by the very solvers it replaces, so it inherits their biases.

The Strongest Case Against This Analysis

Here is the best case against the "GPT moment for simulation" framing. The paper states plainly that the team trains one parameter set per dynamics category: six separately trained models share one architecture, and a single shared-weight model is future work. No foundation model that works out of the box for any scenario exists here yet. What exists is one reusable architecture that still needs retraining per phenomenon, on data generated by the very solvers it claims to replace, capped at 50,000 particles, with an admitted accuracy cost against specialists. Until a single shared-weight model beats a specialized solver on that solver's home turf at production scale, WorldParticle is a unification of engineering effort, not a unification of physics, and most of that is correct; the honest reading is narrower but still important, because building a new simulator collapsed from writing a solver to training a model.

What You Can Do

If you run simulations for a living, treat WorldParticle as a fast first-pass tool, not a certified replacement: sweep a design space with it, rank ten candidate geometries in minutes, then validate the top two with the classical solver you already trust. If you teach physics or engineering, the paper and project page are open access, and the six-category demo shows students that sand, water, cloth, and proteins are the same math wearing different costumes. If you work in drug discovery, watch the protein number, not the headline: a 100x speedup only matters if the error stays low on your molecule, and the paper's protein error is its best result, encouraging but unreplicated. If you build robots, copy the inverse-design demo. Differentiating through a learned simulator to recover friction or stiffness from a target motion works on your own system long before any unified model arrives.

The Bottom Line

One transformer architecture now simulates six kinds of physics, protein motion 100 times faster than the old numerical methods and 14.7 times faster than real time, a 600-frame fluid rollout in 37 seconds, five solver codebases replaced by one design. It is not yet the foundation model its authors dream of: six separately trained weight sets, a 50,000-particle ceiling, and an accuracy trade against the specialists. But the direction is unmistakable: simulation used to mean hand-coding a new solver for every new material, and WorldParticle suggests the next decade means training one architecture and pointing it at whatever moves.

Related

Sources: Wang, Guo, Chen et al., "WorldParticle: Unified World Simulation of Lagrangian Particle Dynamics via Transformer," arXiv:2605.15305; MIT CSAIL / MIT News, "A single AI model simulates the behavior of many substances, from sliding sand to folding proteins" (2026-11-09); UBC Computer Science, same release; HPCWire, "From Sand to Proteins: MIT's WorldParticle Aims for a GPT Moment in Scientific Simulation" (2026-11-11); WorldParticle project page. All real-time factors, wall-clock figures, attention-pair counts, and solver-consolidation counts are the author's own calculations on the cited numbers.