AI models can already write, make images, and create 3D models. Simulating the physical world is harder. A new approach called GeoPT aims to close that gap by giving simulation models a broader way to learn how objects respond to forces.
Why physics is still a hard problem for AI
Text and pixels have become familiar ground for modern AI systems. Physical simulation is different because it asks a model to represent how real objects behave under pressure, motion, wind, water, light, and impact.
Engineers who want reliable simulations need large amounts of physics data. That data is difficult to produce at scale because neural networks depend on information generated by numerical solvers, which calculate physical properties at different points of a 3D shape.
Those solvers are thorough, but slow. The bottleneck matters when engineers want to test designs such as airplanes for safety and aerodynamics, or when they need to understand how robots and vehicles behave in varied environments.
GeoPT, developed by researchers at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) and Tsinghua University, is designed to make that training process broader and more efficient. The approach pre-trains models on virtual mechanical interactions in 3D, helping them build a practical sense of physics before they work on labeled tasks.
How GeoPT learns from synthetic dynamics
The core idea behind GeoPT is synthetic dynamics. In the training setup, small particles interact with complex 3D shapes. Tiny spheres move at different speeds and angles until they reach a point on an object and stop there.
GeoPT studied 1.3 million samples of synthetic dynamics. These samples show the model how particles make contact with objects, giving it a reusable signal about shape, force, and physical response.
The particles do not pass through the object or bounce away. They effectively stop at contact. That simple interaction gives the model a way to connect geometry with physical behavior without requiring the same volume of costly labeled simulation data.
For users, GeoPT is built around 3D objects and force inputs. A user can upload a 3D model, such as a battleship, passenger airplane, or truck, then specify the speed and direction of the force to simulate. The output is a heat map showing how different parts of the object are affected.
That makes the approach relevant to several kinds of physical questions:
- how a car might deform after crashing into a wall
- how light moves around objects
- whether a boat remains afloat in turbulent waves
- how wind and surface pressure affect complex 3D shapes
What the benchmarks showed
The researchers reported strong results on industrial simulation tasks. Across benchmarks, GeoPT outperformed state-of-the-art simulation models, with the same general pattern appearing in multiple settings: it reached peak performance sooner and needed significantly fewer labeled data.
On a dataset involving complex 3D shapes, wind currents, and surface pressure, GeoPT performed better than state-of-the-art models in speed, accuracy, and efficiency. It also showed gains in speed and accuracy when simulating fighter jets responding to wind.
The boat-hull test was another important result. When GeoPT modeled how a hull handled both air and waves, it required 60 percent fewer labeled data to capture both forces and reached peak accuracy four times faster than top baselines.
The system also handled collision and light behavior. It predicted how different types of cars would deform after hitting another object while using less data than state-of-the-art baselines. In another test, it accurately simulated how light would pass through what was essentially a toy rabbit, even though it had not trained on that 3D model or light physics beforehand.
Compared to leading models, GeoPT helped models reach peak performance twice as fast and train on up to 60 percent less data. Co-lead author Haixu Wu, an MIT postdoc and CSAIL researcher, said GeoPT was producing high-fidelity simulations with over 100 million mesh points in seconds.
Why this matters for engineering
The immediate value is practical. If engineers can test vehicles, everyday objects, and robots in simulation with less labeled data, they may be able to explore more designs before moving to physical experiments.
The source examples are concrete: cars and planes, chairs and containers, and robots. The physical elements include wind, water, and collisions. In each case, the promise is not that GeoPT replaces engineering judgment, but that it can help simulation models generalize across more scenarios.
The researchers also frame GeoPT as part of a larger direction: a physics foundation model. In that vision, physics becomes a major training modality for AI, alongside text and pixels.
“We believe physics is the third modality for AI models, after text and pixels,” says MIT PhD student and CSAIL researcher Minghao Guo, a co-lead author on a paper introducing GeoPT.
That framing matters because many AI systems are already strong with language and visual data. Adding physical accuracy could make tools for robotics data, video generation, and simulation more realistic.
What comes next
The researchers describe GeoPT as a preview of the physics world model they are working toward. Their next goal is to scale the system by training on more shapes and simulating more complex physical phenomena.
The possible future uses named by the team include modeling weather patterns, testing different materials, and generating realistic videos. Those are broader targets than the current industrial benchmark results, but they follow from the same core ambition: train AI systems to represent physical behavior more generally.
The paper was written by Wu and Guo with MIT CSAIL colleagues Zongyi Li, Zhiyang (Frank) Dou, Kaiming He, and senior author Wojciech Matusik. Tsinghua University Associate Professor Mingsheng Long was also a co-author. The team presented the paper at the International Conference on Machine Learning in July, and the work was supported, in part, by Neural Modular Physics Twin for Robotics.