MIT and UBC Build WorldParticle, One AI Model for Many Physical Simulations
Researchers at MIT, the University of British Columbia and partners built WorldParticle, one transformer that predicts how particles move across six kinds of physical behaviour. The team reports protein simulations about 100 times faster, but gives no accuracy figures.
By Utopia Data & AI Team6 min read
In brief
- MIT News reported on 9 November 2026 that researchers at MIT, the University of British Columbia and other institutions built WorldParticle, a deep-learning system that simulates many physical behaviours with one architecture.[1]
- It represents materials as particles and predicts each particle's next position and properties. The team tested six phenomena, including deformable objects, granular materials and proteins, and says it beat numerical methods.[1]
- The article reports that protein changes were modelled about 100 times faster than numerical simulators. It names no baselines, gives no accuracy figures and does not state the training data or model size.[1]
What the MIT researchers announced
MIT News reported on 9 November 2026 that researchers from MIT, the University of British Columbia and other institutions developed WorldParticle. It is a single deep-learning system meant to model several kinds of physical behaviour, where today each process usually needs its own hand-coded simulator.[1]
The work will be presented at the ACM SIGGRAPH conference in Asia, and the paper is listed on arXiv under 2605.15305. Peter Yichen Chen, a former MIT postdoc now at UBC, and MIT postdoc Minghao Guo are the co-corresponding authors. Wojciech Matusik of MIT is also named.[1]
This story rests on MIT's own article. The results are the researchers' claims, and we have not read the full paper or run the system. The article states no training data, model size or funding source.[1]
How WorldParticle represents the world
The system breaks an object into thousands or millions of tiny three-dimensional particles. It tracks each particle's position and velocity over time. A neural network takes the particles' positions and physical properties as input and outputs their positions and properties for the next step, which the article describes as one-second increments.[1]
Instead of coding a governing equation for each material, the team says the network learns to solve governing equations in general. That, the researchers say, lets it adapt to materials, forces and environments it did not see in training. The article compares the architecture to ChatGPT's, except that it predicts particle motion rather than words.[1]
External forces, such as an ocean wave hitting a seawall, are separated out before each step. After the network predicts the particle states, the forces are put back to correct the prediction. A merging technique also combines and compresses particles so that computation stays manageable.[1]
What the team reports
The team tested the system on six physical phenomena. The article names deformable objects, granular materials and proteins, and does not name the other three. It says the model was more accurate than numerical simulation methods across the six, without naming those methods or giving error figures.[1]
In one test, a granular object collapsed into a pile. The researchers say some baseline methods left sand particles floating in the air, while WorldParticle did not. For proteins, Guo says the system modelled structure changes over time about 100 times faster than numerical simulators.[1]
The article lists possible uses: soft robot manipulation, RNA evolution, galaxy mergers, oil sloshing against elastic valves in storage tanks, and sand or soil moving under a soft robot. These are goals the researchers describe, not demonstrations reported in the article.[1]
Limits and next steps
Complex objects can contain millions of particles, which quickly becomes too heavy to compute without the merging step. Guo says the team needed a lot of detailed engineering and many tweaks to the architecture before it generalised and scaled. The article does not say how the method behaves on materials outside the six tested.[1]
Chen says the goal is a GPT moment for simulation, meaning a general foundation model for physical systems. Planned work includes adapting the system to more kinds of physical systems and building a general-purpose simulation pipeline for scientific discovery.[1]
Other collaborators come from Georgia Tech, Inria, Meta, the University of Pennsylvania, the University of Utah, UCLA and the University of Hong Kong. The article gives no release plan for code or data.[1]
Why Malaysia should care
Our analysis: MIT names no Malaysian partner. Malaysian universities and engineering firms that rely on physical simulation can read this as an early research signal, not a tool they can buy. The article reports no public code, data or release date, and says nothing about cost. Any local use would need independent testing against the solvers a team already trusts.
Malaysian university engineering and science groups
The work shows one learned model standing in for several hand-built solvers, but only in a research setting.[1]
Practical move: Read the arXiv paper before planning any comparison with your own solvers.
Malaysian engineering and manufacturing firms
No product or release is announced, so there is nothing to adopt yet.[1]
Practical move: Keep existing simulation tools and treat this as research to watch.
Business and technology leaders
A reported speed-up on proteins does not show gains on industrial problems.[1]
Practical move: Do not budget for this on the strength of a news article.
What Malaysians can do now
- Read the arXiv paper 2605.15305 before quoting the 100-times figure.
- Ask which baselines and accuracy measures were used before comparing it with your own tools.
- Watch for a code or data release before planning any trial.
What we still do not know
What we still do not know
- Which numerical methods the system was compared against, and the error figures.
- What data the model trained on and how large it is.
- Whether the code or data will be public.
Sources
- A single AI model simulates the behavior of many substances, from sliding sand to folding proteinsMIT News, 9 November 2026
- File:Avalanche Slope on Dunes (46116909755).jpgWikimedia Commons, 7 March 2018
- File:Sand Avalanche and Ripples (39428290505).jpgWikimedia Commons, 2 November 2017
- File:3D Model of the HIV Capsid (34357117320).jpgWikimedia Commons, 27 April 2017
Facts come from MIT News dated 9 November 2026, read on 22 November. Results and claims are those the researchers reported, and we did not run or audit the system or read the full paper. The section marked as our analysis is Utopia's interpretation. The three photographs are contextual: two US National Park Service images of sand dunes and a NIAID photograph of a 3D-printed model of the HIV capsid, a protein shell. None shows WorldParticle or the MIT team. No generated image is used.
Last fact check: 22 November 2026. Factual correction? devteam.utopia@gmail.com.


