One AI Model Now Simulates Sand, Cloth, Fluids and Folding Proteins as MIT Chases a ‘GPT Moment’ for Simulation

Abdul Mannan 7 minutes ago • 16 min read

The Ray and Maria Stata Center at MIT in Cambridge, Massachusetts, home of the Computer Science and Artificial Intelligence Laboratory (CSAIL), where the WorldParticle research team works. — Photo: Tony Webster / Wikimedia Commons (CC BY 2.0). The Ray and Maria Stata Center at MIT, photographed 5 December 2014 — file photo, not the research itself.

Researchers at the Massachusetts Institute of Technology have built a single artificial intelligence system that can simulate six very different kinds of physical behaviour — from sand sliding into a pile to cloth draping, fluids sloshing and proteins folding — using one unified model where scientists previously needed a separate, hand-built simulator for each phenomenon. The system, called WorldParticle, was announced by MIT News on 9 November 2026, and the team behind it says it is a first step towards a general-purpose “foundation model” for physics itself.

Simulation is one of the quiet workhorses of modern science and engineering. Before a bridge is built, a drug is designed or a robot is deployed, engineers run computer models of how materials will behave. But those models have always been fragmented. A program that predicts how water flows cannot predict how sand pours; a tool that models a bending beam cannot model a folding protein. Each one is built from scratch, on equations derived by hand for that particular material, in that particular situation.

WorldParticle breaks that pattern. According to MIT News, the system treats every material it encounters — whether a dune of sand, a sheet of cloth, a volume of fluid or a protein molecule — as a cloud of thousands or millions of tiny three-dimensional particles, and then predicts where each particle will move next, one step at a time. It does so with the same underlying architecture in every case, learning the dynamics from data rather than being handed the governing equations.

“We want to have a GPT moment for simulation,” said Peter Yichen Chen, a former MIT postdoctoral researcher who is now an assistant professor at the University of British Columbia and a co-corresponding author of the paper describing the technique, according to MIT News. “Every time people want to build a simulator for a particular physical phenomenon, they need to code it up from scratch. Our vision is to have a foundation model that works out of the box for any scenario you want to simulate. This is a first step in that direction.”

The research, published in the peer-reviewed journal ACM Transactions on Graphics, was led by Chen and Minghao Guo, an MIT postdoctoral researcher, with co-lead authors Caoliwen Wang and Siyuan Chen, graduate students at the University of British Columbia, and Wojciech Matusik, an MIT professor of electrical engineering and computer science who leads the Computational Design and Fabrication Group within the Computer Science and Artificial Intelligence Laboratory (CSAIL). Collaborators from Georgia Tech, Inria, Meta, the University of Pennsylvania, the University of Utah, the University of California at Los Angeles and the University of Hong Kong also contributed. The work is due to be presented at the ACM SIGGRAPH Conference and Exhibition on Computer Graphics and Interactive Techniques in Asia.

How one model learns six kinds of physics

The central idea behind WorldParticle is disarmingly simple: stop writing a new simulator for every material, and instead teach one system to watch particles move. The approach is known as a Lagrangian representation, and it is the same way physicists have long described fluids and granular materials — by following the motion of individual parcels of matter rather than by dividing space into a fixed grid.

What is new is that WorldParticle does not need the equation that governs each material. Traditionally, a scientist modelling a fluid starts with the Navier-Stokes equations, a set of partial differential equations describing how velocity, pressure and viscosity interact, and then spends months turning those equations into a numerical solver that a computer can run. Modelling sand requires a different theory, modelling cloth requires yet another, and modelling a protein requires the machinery of molecular dynamics. Each conversion from equation to code is a research project in its own right.

WorldParticle replaces that labour with learning. The system is built on a transformer — the same family of neural network architecture that powers large language models — but instead of predicting the next word in a sentence, it predicts the next position of every particle in a scene. “It is 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,” Chen said, according to MIT News.

The design has a distinctive two-stage rhythm. Before each simulation step, the system separates out the external forces acting on the scene — the kind of sharp impulse produced when an ocean wave slams into a seawall, for example. It first predicts how the particles would move on their own, and then folds those forces back in, correcting its prediction. The paper describes this as a prediction-correction transformer, and the team says the separation is one of the reasons the model can generalise to materials, forces and environments it never saw during training.

Scale is the obvious obstacle. A realistic scene can contain millions of particles, and asking a neural network to track every one of them, frame after frame, quickly becomes computationally unmanageable. To keep the problem tractable, the researchers built in a merging technique that compresses groups of particles while preserving the fine-grained details of their behaviour. “We needed to do a lot of detailed engineering and employ many tweaks to the architecture for it to be truly generalizable and scalable,” Guo said, according to MIT News.

The result is a single architecture that, according to the paper, remains entirely fixed across all six categories of dynamics it was tested on — cloth, elastic solids, Newtonian and non-Newtonian fluids, granular materials and molecular dynamics — with only the training data changing between them. That is the claim that sets WorldParticle apart from earlier learned simulators, which typically specialised in one kind of material or needed structural changes to handle another.

What the tests showed

The team demonstrated the system on six different physical phenomena, and reported that it matched or beat conventional numerical simulation methods on accuracy across the set. The details matter, because simulation is a field where small errors compound: a model that is slightly wrong at every step can drift into nonsense over a long run, producing sand that floats, cloth that explodes or solid objects that pass through one another. The paper reports that established baseline methods — including the graph-network simulator known as GNS, DeepLagrangian and neural operator approaches — showed exactly those artefacts on long rollouts, while WorldParticle held together.

One of the clearest demonstrations involved granular material. When a heap of sand collapses into a pile, WorldParticle captured the collapse convincingly; some of the baseline methods, by contrast, left sand particles floating in mid-air, according to MIT News. It is the kind of failure that looks trivial on a screen and is fatal in engineering, where a simulation that cannot tell a pile from a cloud cannot be trusted to design anything.

The protein results were the most striking on speed. The team reports that WorldParticle modelled how protein structures change over time roughly 100 times faster than numerical simulators, Guo said, according to MIT News. Proteins are chains of amino acids that fold into precise three-dimensional shapes, and that shape determines what a protein does in the body — which is why protein behaviour sits at the centre of drug discovery. A simulator that can run a hundred times faster does not merely save time; it changes which questions are worth asking, because experiments that were too expensive to run become routine.

The paper also shows the system handling finer-grained molecular detail, including the rotation of amino-acid side chains in specific proteins, and demonstrates downstream uses beyond pure prediction: interactive control of a simulated scene, inverse design — working backwards from a desired outcome to the conditions that would produce it — and learning from real-world observations rather than only from synthetic training data.

Why simulation has resisted unification until now

To understand why a single model for six kinds of physics is a notable result, it helps to see how simulation has traditionally been organised. For most of the twentieth century, the field advanced by division. Fluid dynamicists built solvers for fluids. Structural engineers built solvers for solids. Granular physicists built solvers for sand, soil and grain. Molecular dynamicists built solvers for atoms. Each community developed its own mathematics, its own numerical tricks and its own software, and the tools rarely crossed over.

The reason was partly mathematical. The equations that describe a fluid are genuinely different from the equations that describe a rubber band, and a numerical method that is stable for one can be unstable for the other. But the reason was also practical: simulation software is expensive to build and validate, so once a tool worked for its intended purpose, there was little incentive to generalise it. The result is a landscape of thousands of specialised codes, each excellent in its niche and useless outside it.

The last few years have begun to erode that landscape. In 2020, researchers at DeepMind showed that a graph neural network — the system known as GNS — could learn to simulate fluids, sand and deformable solids from data, and the idea of a “learned simulator” entered the mainstream. Physics engines built for robotics and games have also become more general: Nvidia’s Newton engine, released in 2025 and developed with Google DeepMind and Disney Research, aims to give robots a single, open-source physics core to train against. And in biology, DeepMind’s AlphaFold showed in 2020 and 2021 that a learned model could solve a problem — predicting protein structure — that had resisted decades of physics-based attack.

WorldParticle belongs to that wave, but it pushes the idea further in two directions. First, it unifies more kinds of physics under one roof than previous learned simulators attempted, including the notoriously difficult combination of fluids and deformable solids interacting with each other. Second, it does so with a transformer architecture — the same design that scaled language models to general-purpose tools — rather than with a graph network, which is a deliberate bet that the scaling lessons of language AI will transfer to physics.

“If you study deformable objects and fluids, you would need two simulators — one for each physical regime. But if you unify them into one model, maybe you could couple the physics of these two processes and gain insights into new problems?” Guo said, according to MIT News. That coupling is the point. Many of the hardest problems in engineering are not pure fluid problems or pure solid problems, but mixed ones: oil sloshing against a flexible valve inside a storage tank, a soft robot’s gripper deforming as it presses into soil, blood flowing through a bending artery. A simulator that can only handle one side of such a problem is of limited use.

What it could be used for

The researchers sketch a wide range of applications. In robotics, a unified simulator could help engineers design soft robots — machines made of flexible materials that squeeze, bend and grip — by modelling both the robot’s body and the sand, soil or tissue it touches in a single run. In industry, it could model how a sloshing liquid stresses the flexible components of a tank, a problem that currently requires chaining two different simulators together and hoping the join holds.

In biology, the protein results point towards faster exploration of how molecules move and change shape, which matters for understanding disease and designing drugs. The paper also raises the possibility of modelling RNA evolution, and the team says the framework could in principle be extended to astrophysical phenomena — the merging of galaxies is named as an example — where the same particle-based mathematics applies at an unimaginably larger scale.

There is also a longer-term ambition that the researchers state openly: a full, general-purpose simulation pipeline for scientific discovery, in which a researcher describes a physical system and a foundation model simulates it, without the months of solver-building that such a project requires today. “We see the unification of computation and science giving us new insights and tools to discover physical phenomena or explain some unexplainable physical phenomena,” Guo said, according to MIT News. “With a unified model, we could unlock many discoveries that we can’t even imagine.”

The caveats, stated plainly

For all the ambition, the honest limits of the work deserve equal attention — and the researchers and independent coverage of the announcement are clear about them. WorldParticle is a research system, not a production tool. Independent summaries of the announcement note that the ability to generalise to new situations is a goal of the research rather than a guarantee, and that further testing will be needed to establish where the model is accurate enough to support practical decisions.

One important qualification concerns how unified the model truly is today. Although the architecture stays fixed across phenomena, the system is trained separately for each category of dynamics — one parameter set per category — so a single shared model that handles everything at once remains future work, according to independent analysis of the announcement. The team also trades some per-phenomenon accuracy against specialist solvers: a bespoke simulator built for one material can still beat a generalist on its home turf. That is the familiar price of generality, and it is one the team will need to keep paying down.

The training data itself carries a caveat. Like all learned simulators, WorldParticle learns from examples, and its behaviour outside the range of those examples is not guaranteed. A model trained on gentle fluid flows may misbehave on violent ones; a model trained on one kind of sand may stumble on another. The paper’s tests on unseen materials and forces are encouraging, but they are tests on a research benchmark, not a warranty.

None of that diminishes the result. The history of simulation is a history of generalisation arriving in steps: first one material, then a family of materials, then — if the bet pays off — everything. WorldParticle is a serious step in that sequence, published in a peer-reviewed journal and demonstrated against strong baselines. Whether it becomes the foundation model its authors hope for will depend on what happens next: whether the code is released, whether other laboratories reproduce and extend the results, and whether the system survives contact with the messy, unlabelled data of the real world.

Analysis: Why It Matters

The end of the bespoke simulator

The most important idea in this announcement is not a new equation or a faster chip. It is an organisational idea: that simulation, like language, might be a single skill rather than a thousand separate ones. For decades, the field assumed that each physical phenomenon needed its own mathematics, its own code and its own experts. WorldParticle challenges that assumption directly, and if the challenge holds, the economics of science change with it. A small laboratory that cannot afford to build a custom solver for its problem could instead borrow a foundation model and get to work — the same democratising effect that large language models had on software, applied to physics.

A foundation model for physics, not just for words

The phrase “GPT moment” is doing deliberate work in this story. The team is not claiming to have built the finished article; it is claiming that simulation has reached the point language reached a few years ago, when a single architecture trained at scale turned out to be useful for tasks nobody had designed it for. The analogy is imperfect — language has vastly more training data than physics, and a wrong word is cheaper than a wrong bridge — but the direction of travel is the same. The researchers’ bet is that the transformer, which learned the statistical structure of text, can also learn the dynamical structure of matter. The early evidence, across six kinds of physics, suggests the bet is at least worth taking seriously.

Speed is a scientific instrument, not just a convenience

The hundred-fold speedup on proteins deserves more attention than it has received. In science, speed is not merely a matter of waiting less; it changes what can be studied at all. A simulation that takes a week can only be run a handful of times, so researchers study a handful of cases. A simulation that takes an hour can be run hundreds of times, so researchers can explore, sweep parameters, test hypotheses and stumble onto things they were not looking for. If WorldParticle’s speed holds up on real biological problems, it could do for protein dynamics what AlphaFold did for protein structure: turn a scarce resource into a plentiful one, and let the questions get bigger.

The coupling dividend

The most valuable problems in engineering are mixed problems, and they are exactly the problems that fragmented simulation handles worst. A heart valve is a fluid problem and a solid problem at once. A landslide is a granular problem and a fluid problem at once. A soft robot is a solid problem and a contact problem at once. Today, engineers attack such problems by coupling separate simulators, a fragile process in which errors at the boundary can swamp the physics. A single model that learns the coupled dynamics directly, as WorldParticle attempts, removes the boundary altogether. That is a quieter claim than the “GPT moment” framing, and it may prove the more durable one.

What to watch next

The story is not finished; in several ways it has barely started. The paper’s presentation at SIGGRAPH Asia will put the work before the graphics and simulation community, where the toughest questions will come from people who build simulators for a living. The release of code and trained models — standard practice in machine learning, but not yet confirmed here — will determine whether other laboratories can test the claims independently. And the real test will be generalisation: whether a model trained on laboratory benchmarks can handle the unlabelled, noisy, out-of-distribution data of real engineering and biology. The team says it wants to build towards a full general-purpose simulation pipeline for scientific discovery. The measure of that ambition will be whether, a year from now, other researchers are using WorldParticle to answer questions they could not answer before — or whether it remains an elegant demonstration. The odds look better than they have for any previous attempt, and that, for a field as conservative as simulation, is news in itself.

What to Watch Next

The SIGGRAPH Asia presentation, where the graphics and simulation community will scrutinise the results and the baselines.
Whether the team releases code and trained models, which would let independent laboratories reproduce and extend the work.
Tests of generalisation to unseen materials, forces and environments beyond the six benchmark categories.
Applications in soft robotics, where a unified model of deforming bodies and granular contact could change how robots are designed and trained.
Whether the hundred-fold protein speedup holds on real drug-discovery workloads, not only on benchmark proteins.
Follow-up work on a single shared model across all phenomena, rather than one parameter set per category.
Whether rival approaches — graph-network simulators, physics engines such as Nvidia’s Newton, and hybrid methods — converge on the same transformer-based design or take a different path.