GeoPT_Pretrained_Models

GeoPT: Scaling Physics Simulation via Lifted Geometric Pre-Training

Project Page | Paper | GitHub

GeoPT is a unified pre-trained model for general physics simulation based on lifted geometric pre-training. It bridges the geometry-physics gap by augmenting geometry with synthetic dynamics, enabling dynamics-aware self-supervision without the need for expensive physics labels.

Key Features

• Data Efficiency: Reduces labeled training data requirements by 20–60% across diverse physics simulation tasks.
• Scalable Self-Supervision: Generates millions of training samples quickly, significantly faster than traditional physics supervision.
• Strong Scaling: Performance consistently improves with larger models and more training data.
• Generalization: Generalizes across diverse physical systems (fluid and solid mechanics) by reconfiguring the dynamics condition as a "prompt".

Usage

Concrete usage instructions, including environment setup and scripts for fine-tuning GeoPT on various industrial-fidelity benchmarks (such as AirCraft, Cars, and Ships), can be found in the official GitHub repository.

Citation

If you find this repo useful, please cite the paper:

@article{wu2026GeoPT,
  title={GeoPT: Scaling Physics Simulation via Lifted Geometric Pre-Training},
  author={Haixu Wu, Minghao Guo, Zongyi Li, Zhiyang Dou, Mingsheng Long, Kaiming He, Wojciech Matusik},
  booktitle={arXiv preprint arXiv:2602.20399},
  year={2026}
}

Contact

If you have any questions, please contact Haixu Wu (wuhaixu98@gmail.com) and Minghao Guo (guomh2014@gmail.com).

Acknowledgement

We appreciate the following GitHub repositories for their valuable codebase or datasets: Transolver and Neural-Solver-Library.