GenCP: Towards Generative Modeling Paradigm of Coupled Physics
Tianrun Gao, Haoren Zheng, Wenhao Deng, Haodong Feng, Tao Zhang, Ruiqi Feng, Qianyi Chen, Tailin Wu
TL;DR
GenCP introduces a novel generative modeling framework for simulating coupled physical systems by integrating probability density evolution with iterative multiphysics coupling, enabling efficient and accurate simulations from decoupled data.
Contribution
It formulates coupled-physics modeling as a probability modeling problem and uses operator-splitting theory to provide error guarantees, advancing simulation capabilities for complex systems.
Findings
Demonstrates superior performance on multi-physics scenarios
Enables training on decoupled simulation data
Provides error controllability guarantees
Abstract
Real-world physical systems are inherently complex, often involving the coupling of multiple physics, making their simulation both highly valuable and challenging. Many mainstream approaches face challenges when dealing with decoupled data. Besides, they also suffer from low efficiency and fidelity in strongly coupled spatio-temporal physical systems. Here we propose GenCP, a novel and elegant generative paradigm for coupled multiphysics simulation. By formulating coupled-physics modeling as a probability modeling problem, our key innovation is to integrate probability density evolution in generative modeling with iterative multiphysics coupling, thereby enabling training on data from decoupled simulation and inferring coupled physics during sampling. We also utilize operator-splitting theory in the space of probability evolution to establish error controllability guarantees for this…
Peer Reviews
Decision·ICLR 2026 Poster
I think the proposed method is easy to follow and the result is good. Also, I think the problem the authors are trying to solve is meaningful.
1. I think the experiments are not enough. It all concentrates on fluid dynamics, but it is better to combine fluid and rigid objects. It is convenient to treat the whole fluid system as a domain. 2. Also for the experiments part, I feel like the submission should add a baseline: training a single v(x_t, y_t) for all variables (coupled data), to prove the decouple the data is beneficial.
Combining probability density evolution with iterative multiphysics coupling in generative modeling is a novel approach, and the problem setting is very clear.
The theoretical part of this paper is excellent, but I think the experimental part could be scaled up further.
- According to the Reviewer, the general idea presented in the paper of integrating operator-splitting theory with generative flow models is novel. Using decoupled training and coupled inference later to reduce the cost of training data could potentially be very impactful as many AI4Science ML applications deal with the issue of spending significant compute on training data generation. - The authors present theoretical guarantees for their framework. The method is not just a heuristic but direc
Despite the core idea being novel, the Reviewer is unable to recommend the paper for acceptance at ICLR due to the following major issues: - One of the central motivation of the paper is that the training data generation using decoupled systems only is cheaper but this was unfortunately never quantified. The reviewer recommends evaluating the cost of training data generation for coupled and decoupled generation. - The comparisons to baselines should also include frameworks trained on the fully c
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsGenerative Adversarial Networks and Image Synthesis · Model Reduction and Neural Networks · Tensor decomposition and applications