Artificial intelligence is opening up new possibilities for scientific simulations, helping researchers study how different materials behave under changing conditions. From flowing fluids and shifting sand to the complex movements of proteins, scientists are exploring whether AI can model a wide range of physical systems using a common approach.
A research team involving the Massachusetts Institute of Technology (MIT) and the University of British Columbia (UBC) has developed a system called WorldParticle. Announced by MIT on November 9, 2026, the system uses a transformer-based architecture to predict how particles move and interact over time. Researchers tested it across six categories of physical dynamics, including fluids, cloth, elastic materials, granular matter and molecular motion.
AI is also being explored for a wider range of scientific applications, as discussed in our coverage of AI scientific discoveries
Traditionally, different physical systems often require specialised simulation methods. WorldParticle takes a different approach by using the same underlying AI architecture across these areas, potentially making it easier to develop neural simulators for different types of materials.
There is, however, an important detail behind the research. WorldParticle is not yet a single AI model trained to simulate every category. The researchers trained separate models for each category using the same architecture. Developing one model with shared weights across all categories remains a goal for future research.
What Is WorldParticle AI?
WorldParticle is a learning-based physics simulation system developed by researchers from MIT, UBC and several other institutions.
Traditional physics simulators rely on mathematical equations and numerical methods to predict how materials and objects behave. A simulator designed to model flowing water, for example, may use different mathematical techniques from one built to simulate cloth deformation or protein movement.
WorldParticle takes a different approach. Rather than developing a completely new neural network architecture for every physical process, the researchers created a transformer-based architecture that can handle several types of particle dynamics.
The system represents physical objects and materials as collections of small three-dimensional points called particles. Each particle carries information such as its position, velocity and relevant physical properties.
The AI uses this information to predict how the particles will move during the next simulation step.
According to the original research paper, the architecture was tested across six categories of physical dynamics:
- Newtonian fluids, such as water-like liquids.
- Non-Newtonian fluids, whose flow behaviour changes under different conditions.
- Cloth and other deformable materials.
- Elastic solids.
- Granular materials, including sand.
- Molecular dynamics involving proteins.
The researchers used the same general architecture across these categories instead of designing a completely different network structure for every physical process.
How Does WorldParticle Work?
To understand WorldParticle, it helps to look at how a conventional physical simulation works.
A physical simulator must estimate how forces, material properties and interactions affect an object over time. These calculations can become difficult when thousands or millions of particles interact with one another.
WorldParticle uses a prediction-and-correction approach to handle this problem.
1. Representing Materials as Particles
The first step is to represent a material or object as a collection of particles.
Each particle has a position and velocity, along with physical information relevant to the material being simulated. For example, a fluid particle may have properties associated with its flow, while a protein particle represents an atom in the molecular system.
This shared representation gives the architecture a consistent way to process different physical systems.
2. Predicting the Effect of External Forces
Before the AI corrects particle interactions, a mathematical prediction step accounts for known external forces.
These can include gravity, applied forces or other conditions affecting the simulated material.
The prediction produces an intermediate state showing how particles would move under those external forces before the model accounts for their interactions with one another.
3. Correcting Particle Movement With AI
A learned neural network then predicts adjustments to particle positions and velocities.
This correction step helps account for interactions that are not fully represented by the initial external-force calculation.
The architecture uses local particle information and a compressed representation called super tokens to communicate information across the simulated system.
This combination is important because physical behaviour can depend on both nearby interactions and effects that travel across larger portions of an object or material.
Why Does WorldParticle Use Super Tokens?
Simulating millions of particles can require substantial computing resources. One reason is that the model needs to understand how particles influence one another.
WorldParticle uses a token-merging process to reduce the amount of information handled by later stages of its transformer architecture.
The system first creates representations of individual particles and their local interactions. It then combines related particle tokens into smaller groups called super tokens.
These super tokens provide a compressed representation of the larger particle system. The decoder uses that information to produce corrections for individual particles.
This approach allows the model to capture local details while also communicating information across the wider simulation.
However, token merging does not eliminate every computing limitation. The researchers note that the first encoder layer still performs attention at full particle resolution, which contributes to the system’s memory requirements.
What Did the Researchers Test?
The research team evaluated WorldParticle across six physical dynamics categories using data generated by established simulation tools.
The experiments included different materials, initial conditions, physical parameters and boundary configurations. The researchers also tested whether the models could handle configurations that were not included in their training sequences.
Some of the examples included:
- Fluids: Predicting how liquid moves around obstacles and through containers.
- Cloth: Simulating fabric falling onto objects and responding to collisions.
- Sand: Predicting how granular material collapses into piles under gravity.
- Non-Newtonian fluids: Modelling materials whose flow behaviour changes with their physical conditions.
- Elastic solids: Predicting how deformable objects move and interact with surfaces.
- Proteins: Modelling changes in molecular structures and side-chain movement.
The paper reports that WorldParticle produced stable predictions across several unseen configurations within its tested categories.
For example, the researchers demonstrated cloth simulations that remained stable for longer than the training rollout horizon. They also tested fluid simulations with different initial arrangements and boundary conditions.
These results suggest that the architecture can generalise beyond the exact examples used during training. They do not establish that it can accurately simulate every possible material or physical process.
An Important Catch: One Architecture Does Not Mean One Universal AI Model
Some descriptions of WorldParticle may give readers the impression that a single trained AI model has already learned to simulate every type of material.
That is not what the research paper demonstrates.
The researchers explicitly state that they currently train a separate model for each dynamics category. Each model uses its own learned parameters, even though the underlying transformer architecture remains the same.
In practical terms, the fluid model and the protein model are separately trained versions of the same general design. They are not one identical set of learned weights handling every category.
The researchers describe training one model with shared weights across all categories as a future direction. Achieving that would move the project from a unified architecture to a genuinely unified model.
This distinction is important because the current research demonstrates that one architecture can support several kinds of physical simulation. It has not yet demonstrated one universal model that handles all those categories together.
What Are the Current Limitations of WorldParticle?
WorldParticle shows promising results, but the researchers also identify limitations that need to be addressed.
Computing and Memory Requirements
One major limitation is memory usage.
According to the paper, the current architecture supports approximately 50,000 particles on a single NVIDIA A100 GPU with 80GB of memory.
This limit is connected to the full-resolution attention operation in the first encoder layer. Larger simulations could require more memory-efficient attention methods or other architectural improvements.
Separate Training Is Still Required
The system also requires a separately trained model for each physical dynamics category.
Training one model with shared weights across all categories remains future work. The researchers note that combining the training data and maintaining accuracy across different categories could create additional challenges.
Accuracy Can Differ From Specialised Simulators
A general-purpose architecture also involves trade-offs.
The authors state that their approach sacrifices some per-category accuracy compared with specialised physical simulation methods.
The paper reports improvements over existing neural simulators in its evaluated comparisons, but that should not be confused with proving that WorldParticle is more accurate than every established physics solver.
This distinction matters because specialised numerical methods remain important for applications where high physical accuracy is essential.
More Physical Systems Need to Be Explored
The researchers want to extend the architecture to additional physical systems.
Their future directions include areas such as soft robotics, RNA-related processes and astrophysical simulations. These are potential research applications, not capabilities already demonstrated across all those fields.
Could WorldParticle Help With Scientific Research?
The potential applications extend beyond producing visual simulations.
One area is soft robotics. Researchers could use particle-based models to explore how flexible objects respond when forces are applied, potentially helping with the design and control of soft robotic systems.
Another area is material science. More flexible simulation approaches could help researchers investigate how different materials behave under changing conditions.
Protein simulation is also an interesting direction. Understanding how molecular structures change over time is important in computational biology and related research.
Related research into wetware computing and human brain cells explores another approach to combining biological systems with modern computing technologies.
The paper additionally demonstrates interactive control, inverse design and predictions based on real-world manipulation data.
In the real-world data experiment, the researchers used particle trajectories extracted from three manipulation tasks, including lifting cloth, stretching a toy and folding a rope. The results showed qualitative agreement between predicted and observed motion.
These experiments are early demonstrations of how learned simulation systems might connect computational models with physical observations. They do not mean WorldParticle is already a validated tool for medical research, drug discovery or industrial deployment.
What Makes WorldParticle Different From Traditional AI Simulation?
The main contribution is not simply that AI can generate realistic-looking movement.
WorldParticle is designed to predict particle-level physical dynamics through a shared transformer architecture. It combines a mathematical prediction step with a learned correction process.
Traditional simulation systems often rely on specialised numerical methods for particular physical processes. WorldParticle explores whether a common learned architecture can reduce some of the engineering needed to build separate neural simulators.
However, specialised solvers still have an important role, especially when accuracy and physical reliability are critical.
The research therefore represents progress toward more flexible scientific simulation rather than a replacement for established physics software.
What Happens Next for WorldParticle?
The research team wants to extend WorldParticle to more physical systems and work toward a broader simulation framework for scientific discovery.
One of the most important future goals is training a single model with shared weights across different dynamics categories. This would be a major step beyond the current setup, where each category has its own trained model.
The researchers will also need to address memory requirements and investigate how well the approach performs across additional physical conditions and practical applications.
The research paper was submitted to arXiv in May 2026 and last revised on May 20, 2026. MIT published its research announcement on November 9, 2026, and stated that the work would be presented at the ACM SIGGRAPH Asia conference.
For now, WorldParticle is best understood as a research-stage approach to more flexible physics simulation. Its significance lies in demonstrating how one transformer architecture can be adapted to several physical systems, while leaving the larger goal of a genuinely universal simulation model for future work.
Frequently Asked Questions
What is WorldParticle AI?
WorldParticle is a transformer-based AI system developed by researchers from institutions including MIT and the University of British Columbia. It predicts how particles move over time to simulate different physical processes, including fluids, cloth, sand, elastic materials and protein dynamics.
Can WorldParticle simulate sand, water and proteins?
Yes. The research demonstrates simulations across six categories of physical dynamics, including granular materials such as sand, Newtonian and non-Newtonian fluids, cloth, elastic solids and molecular dynamics involving proteins.
Is WorldParticle one universal AI model?
No. WorldParticle currently uses separate trained models for different dynamics categories. Although these models share the same underlying architecture, a single model with shared weights across all categories remains a goal for future research.
How does WorldParticle simulate physical materials?
WorldParticle represents materials as collections of particles and predicts how they move over time. Its transformer-based neural network processes particle interactions to update their predicted positions and velocities during each simulation step.
How many particles can WorldParticle simulate?
The researchers report that WorldParticle can handle approximately 50,000 particles on a single NVIDIA A100 GPU with 80GB of memory under the tested configuration.
Is WorldParticle better than traditional physics simulators?
The research reports better results than selected existing neural simulators in the evaluated comparisons. However, the authors acknowledge some loss of accuracy compared with specialised simulation methods designed for individual physical systems. The study does not establish that WorldParticle outperforms every traditional physics solver.
Can WorldParticle be used for drug discovery?
The research includes simulations involving protein dynamics, but it does not establish WorldParticle as a validated drug-discovery tool. Its potential applications in molecular research will require further investigation and validation.
Sources
- MIT News – A Single AI Model Simulates the Behavior of Many Substances, From Sliding Sand to Folding Proteins
- arXiv – WorldParticle: Unified World Simulation of Lagrangian Particle Dynamics via Transformer
- WorldParticle – Official Research Project Page
