Researchers at MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) and Tsinghua University have created a new pre-training method called GeoPT that helps AI models simulate physics more efficiently and accurately, reaching peak performance twice as fast while using up to 60 percent less data than current leading models. The approach could let engineers test vehicles, robots, and everyday objects against wind, water, and collisions before building physical prototypes.
Most AI models handle text and images well but struggle with physics simulation because they need enormous amounts of specialized data. Current methods rely on "numerical solvers" that calculate physical properties across 3D shapes point by point. This process is thorough but slow, limiting how many design scenarios engineers can test.
GeoPT works differently. It learns physics from "synthetic dynamics": 1.3 million simulated interactions where tiny virtual particles move at various speeds and angles until they stick to a 3D shape. This gives the model a basic sense of physical behavior before it trains on labeled data. Users upload 3D models of objects like trucks or ships, specify wind direction and speed, and get a heat map showing stress points across the surface.
Industry tests show strong results
On benchmarks involving complex 3D shapes responding to wind and surface pressure, GeoPT outperformed state-of-the-art simulation models in speed, accuracy, and efficiency. When testing fighter jets under wind conditions, the system reached similar results. For boat hulls facing both air and waves, GeoPT required 60 percent fewer labeled data points and achieved peak accuracy four times faster than leading baselines.
The system also correctly predicted how 3D vehicles would deform after collisions, and simulated light passing through a toy rabbit shape despite never training on that model or on light physics. "If your model performs well on industrial benchmarks, that means it can solve the hardest physics tasks," says co-lead author Haixu Wu, an MIT postdoc and CSAIL researcher. "GeoPT was making high-fidelity simulations with over 100 million mesh points in seconds."
Foundation model potential
The researchers see GeoPT as an early step toward a physics foundation model - a general-purpose backbone trained on large datasets that can handle diverse tasks. "We believe physics is the third modality for AI models, after text and pixels," says MIT PhD student and CSAIL researcher Minghao Guo, co-lead author. "Our general-purpose model has the versatility to help build a world model for physics."
The team plans to scale the system with more shapes and simulate more complex phenomena, including weather patterns, material behavior, and realistic video generation for AI for Science & Research applications. Their work was supported by the Neural Modular Physics Twin for Robotics project and presented at the International Conference on Machine Learning in July.
Why this matters for science and research professionals
For researchers and engineers working with physical simulations, GeoPT demonstrates that pre-training on synthetic physics data can dramatically reduce the time and cost of simulation, producing accurate results from less specialized data. The approach suggests a shift from today's task-specific simulation models toward general physics foundation models that can handle multiple domains - aerodynamics, fluid dynamics, structural deformation, and optics - without separate training efforts.
Your membership also unlocks:
