One AI Model Can Now Simulate Sand, Cloth, Water and Folding Proteins — Meet WorldParticle

Imagine a single computer programme that can show you how sand slides down a dune, how cloth folds, how water splashes, and how a protein folds itself into shape inside your body. Until now, scientists needed a different simulator for each of those jobs. A team from MIT and the University of British Columbia has built one AI model that can do all six — and it does some of them faster and more accurately than the specialist tools that came before it.

The system is called WorldParticle. It was announced on 9 November 2026, and the research behind it is published in the journal ACM Transactions on Graphics. It will also be presented at the ACM SIGGRAPH Asia conference, one of the biggest gatherings in computer graphics.

The researchers describe it as a first step towards a “foundation model” for simulation — one tool that works out of the box for almost any physical process, in the same way that a chatbot can answer questions on almost any topic.

What Is WorldParticle?

WorldParticle is a deep-learning system that simulates how things move and change. It was built by researchers at MIT, the University of British Columbia (UBC), 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.

The project was led by Peter Yichen Chen, a former MIT postdoctoral researcher who is now an assistant professor at UBC, and Minghao Guo, a postdoctoral researcher at MIT. They worked with UBC graduate students Caoliwen Wang and Siyuan Chen, and with Wojciech Matusik, an MIT professor who leads the Computational Design and Fabrication Group at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL).

The model can simulate six different kinds of physical behaviour:

• Cloth — how fabric bends, folds and drapes
• Elastic solids — soft, bendable objects that spring back into shape
• Newtonian fluids — ordinary liquids such as water
• Non-Newtonian fluids — liquids that behave strangely, such as mixtures that thicken when you stir them
• Granular materials — sand, gravel and other collections of small grains
• Molecular dynamics — the movement of molecules, including proteins

Normally, each of those needs its own simulator, built from scratch, with its own equations and its own software. WorldParticle replaces that patchwork with a single architecture.

“We want to have a GPT moment for simulation,” says Chen. “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.”

How Does It Work?

The trick is that WorldParticle treats everything as particles.

Everything becomes tiny dots

Whether it is looking at a dune of sand or a single protein molecule, WorldParticle breaks the object down into thousands — or even millions — of tiny three-dimensional dots, called particles. This way of describing the world is known as a Lagrangian representation. The model tracks the position and speed of each particle over space and time, and from that it can work out how the whole object moves and changes shape.

It predicts one step at a time

Once an object has been turned into particles, a neural network takes the particles’ positions and physical properties as its input, and predicts where they will be in the next frame of the simulation — one second at a time.

The architecture is a transformer, the same family of AI model that powers ChatGPT. But instead of predicting one word at a time to write a piece of text, WorldParticle predicts how particles will move, one step at a time.

“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 says.

It separates forces, then puts them back

Before each step, the system separates out the outside forces acting on the object — for example, the push of an ocean wave slamming into a seawall. It first predicts where the particles would go without those forces, then folds the forces back into the simulation, correcting its predictions based on how the forces change the object’s behaviour.

This prediction-and-correction design is one reason the model can handle materials, forces and environments it never saw during training. It learns the underlying rules of movement, rather than memorising one fixed situation.

Handling millions of particles

A real object can contain millions of particles, and tracking every one of them would quickly overwhelm even a powerful computer. To keep the work manageable, the researchers use a merging technique that compresses groups of particles together while still capturing the fine details of how they behave.

“We needed to do a lot of detailed engineering and employ many tweaks to the architecture for it to be truly generalizable and scalable,” says Guo.

What Can It Already Do?

The team tested WorldParticle on all six types of physical behaviour, and in their tests it matched or beat the standard numerical simulation methods on accuracy.

Two results stand out.

First, granular materials. When the researchers simulated a pile of sand collapsing, WorldParticle captured the way the grains fell and settled into a pile. Some of the standard comparison methods, by contrast, left sand particles floating in mid-air — an obvious mistake that WorldParticle avoided.

Second, proteins. The model simulated how protein structures change over time about 100 times faster than numerical simulators, according to Guo. Proteins are the tiny machines that do most of the work inside living cells, and the way they fold and change shape decides what they can do. Being able to model that process quickly could matter a great deal for biology and medicine.

The system also showed it could generalise — that is, it worked on new materials, new starting conditions, new boundaries and new forces that it had not been trained on.

Why Does This Matter?

Simulation is one of the quiet workhorses of modern science and engineering. It is how engineers test a bridge before it is built, how car makers test a crash without crashing a real car, and how scientists study processes too small, too large, too fast or too dangerous to watch directly.

Today, building a simulator is slow, specialist work. A researcher who wants to model a new material usually has to write the governing equations for that material, turn those equations into a form a computer can solve, and repeat the whole process for the next material. WorldParticle points to a different future: one model that learns the general rules of physics and applies them anywhere.

The researchers say the framework could be applied to many areas, including:

• Soft robots — machines made from flexible materials, where engineers need to predict how the robot will bend and grip
• RNA and protein research — modelling the molecules of life, which could help with understanding disease and developing new medicines
• Astrophysics — simulating huge physical events such as the merging of galaxies
• New materials — discovering materials with useful properties by testing them in simulation first

“In the long term, the researchers hope this work helps scientists discover new materials and explore physical processes they would be unable to effectively study with existing simulation methods,” MIT News reported.

Guo adds: “We see the unification of computation and science giving us new insights and tools to discover physical phenomena or explain some unexplainable physical phenomena. With a unified model, we could unlock many discoveries that we can’t even imagine.”

What It Cannot Do Yet

WorldParticle is a research result, not a finished product, and it has clear limits.

First, it trades a little accuracy for generality. A simulator built for one specific job can still beat it on that one job. WorldParticle’s strength is doing many jobs well with one system, not being the best at every single one.

Second, the model is not yet fully unified in practice. A separate set of parameters is trained for each category of behaviour — so one trained model handles sand, another handles cloth, and so on. A single shared model that covers everything at once is still future work.

Third, scale has limits. The paper reports the system currently supports up to around 50,000 particles on a single high-end graphics processor (an 80GB A100 GPU). Real-world objects can contain far more particles than that, so bigger simulations will need more computing power or further engineering.

These are normal limits for a first version of a new idea. The researchers say their next steps are to adapt the system to more physical systems and to build towards a full, general-purpose simulation pipeline for scientific discovery.

The Bigger Picture: AI Learns the Rules of the Physical World

WorldParticle is part of a wider trend in AI research known as “world models” — systems that learn the rules of the physical world so they can predict how things will change over time. The same idea is being explored for robots, self-driving cars and video generation.

The difference here is the target: not pictures or words, but physics itself. If a single model can learn how sand, water, cloth and proteins move, then the same approach could eventually be used for almost anything that moves or changes — from weather and oceans to living cells and distant galaxies.

That is why the researchers talk about a “GPT moment” for simulation. ChatGPT changed what people expected from language software: one model, many tasks, no need to build a new tool for each question. WorldParticle is an attempt to do the same for the physical world.

It is early, and the model has real limits. But if the idea works at larger scales, simulation could become as easy to use as a chatbot — and that would change how scientists discover things.

Frequently Asked Questions

What is WorldParticle?

WorldParticle is an AI system built by researchers at MIT, the University of British Columbia and partner institutions. It is a single deep-learning model that can simulate six types of physical behaviour — cloth, elastic solids, ordinary and unusual fluids, granular materials such as sand, and molecules such as proteins — by treating everything as tiny particles and predicting how they move.

How is it different from normal simulators?

Normal simulators are built for one material or process at a time, using hand-written equations. WorldParticle uses one neural network architecture for all six types of behaviour. It learns the rules of movement from data instead of being programmed with a separate equation for each material.

How fast is it?

In the researchers’ tests, WorldParticle modelled how protein structures change over time about 100 times faster than standard numerical simulators. It also matched or beat those methods on accuracy across the six types of behaviour tested.

What could it be used for?

The researchers point to soft robots, RNA and protein research, astrophysical events such as merging galaxies, and discovering new materials. In the long term, they hope it helps scientists study physical processes that existing methods cannot handle well.

Is WorldParticle available to use?

No. It is a research project described in a paper published in ACM Transactions on Graphics and presented at SIGGRAPH Asia. The researchers plan to extend it towards a general-purpose simulation system, but there is no public product yet.

Conclusion

WorldParticle is a small step with a big idea behind it: that one AI model can learn the rules of the physical world well enough to simulate almost anything. It already handles sand, cloth, water and folding proteins in a single system, beats specialist tools on some tests, and models proteins about 100 times faster than the standard methods.

Its limits are real — it is not yet a single shared model, and it gives up a little accuracy for generality. But as a proof that unified simulation is possible, it is a striking result. If the researchers can scale it up, the way scientists test ideas about the physical world could change — from writing a new simulator for every problem, to asking one model to simulate anything.

Sources

• MIT News, “A single AI model simulates the behavior of many substances, from sliding sand to folding proteins”, 9 November 2026: https://news.mit.edu/2026/single-ai-model-simulates-many-substances-behavior-1109
• Wang, Guo, Chen et al., “WorldParticle: Unified World Simulation of Lagrangian Particle Dynamics via Transformer”, ACM Transactions on Graphics (arXiv:2605.15305): https://arxiv.org/pdf/2605.15305
• Technology.org, “A single AI model simulates the behavior of many substances, from sliding sand to folding proteins”, 9 November 2026: https://www.technology.org/2026/11/09/a-single-ai-model-simulates-the-behavior-of-many-substances-from-sliding-sand-to-folding-proteins/

This article is for general information and education only. It is not legal or financial advice.