M2PDE: Compositional Generative Multiphysics and Multi-component PDE Simulation

Authors: Tao Zhang, Zhenhai Liu, Feipeng Qi, Yongjun Jiao, Tailin Wu

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate M2PDE on two multiphysics tasks reaction-diffusion and nuclear thermal coupling where it achieves more accurate predictions than surrogate models in challenging scenarios. We then apply it to a multi-component prismatic fuel element problem, demonstrating that M2PDE scales from single-component training to a 64-component structure and outperforms existing domain-decomposition and graph-based approaches.
Researcher Affiliation Academia 1State Key Laboratory of Advanced Nuclear Energy Technology, Nuclear Power Institute of China, China 2Department of Artificial Intelligence, Westlake University, China. Correspondence to: Tailin Wu <EMAIL>, Yongjun Jiao <EMAIL>.
Pseudocode Yes Algorithm 1 Algorithm for multiphysics simulation by M2PDE. Algorithm 2 Algorithm for multi-component simulation by M2PDE
Open Source Code Yes The code is at github.com/AI4Science-Westlake U/M2PDE.
Open Datasets Yes (2) We create and open-source benchmark datasets for both multiphysics and multi-component PDE simulations, providing a valuable resource for future research. ... As another contribution to the community, we also open-source the data at here to facilitate future method development of multiphysics and multi-component PDE simulations.
Dataset Splits Yes The training data consists of decoupled data, where other physical processes are assumed and treated as inputs to solve the equations governing the current physical process ... The validation data similarly consists of decoupled data not used during training. The test data consists of coupled solutions obtained using fully coupled algorithms.
Hardware Specification No No specific hardware details (like GPU/CPU models or cloud resources) are mentioned in the paper. It only vaguely refers to 'GPU computing' in the efficiency analysis without specifying the type or model of GPU used.
Software Dependencies No The paper mentions several software components like 'Python', 'MOOSE', 'U-Net', 'FNO', 'Geo-FNO', and 'Transolver', but it does not provide specific version numbers for any of these, which is required for reproducible software dependencies.
Experiment Setup Yes The coefficients µu, µv, α, and β of Eq. 18 are set to 0.01, 0.05, 0.1, and 0.25, respectively. The spatial mesh consisted of nx = 20 points, the time step is adaptively controlled by the algorithm, but only outputs the results of 10 time steps. ... The diffusion step of the diffusion model is set to 250.