GenCP : Towards Generative Modeling Paradigm of Coupled physics

Tianrun Gao Haoren Zheng Wenhao Deng Haodong Feng, Tao Zhang Ruiqi Feng Qianyi Chen Tailin Wu · 2026

Read GenCP : Towards Generative Modeling Paradigm of Coupled physics by Tianrun Gao Haoren Zheng Wenhao Deng Haodong Feng, Tao Zhang Ruiqi Feng Qianyi Chen Tailin Wu (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-19541.txt.

Introduction

Tianrun Gao and seven colleagues tackle the simulation of systems where several kinds of physics interact. Heat, flow, and stress often act together in the real world. Modeling them jointly is valuable but hard. Many current methods train on data where the physics were split apart, which does not match real coupled systems.

The team's approach, GenCP, is a generative modeling framework for coupled physics. It aims to handle strong links between the different processes in space and time. Split-data methods tend to lose both speed and accuracy in that setting. A model that learns the coupling directly could serve engineering design better. The work speaks to fields from materials to climate, wherever physics refuses to stay separate.

Catalog record

CREATORTianrun Gao Haoren Zheng Wenhao Deng Haodong Feng, Tao Zhang Ruiqi Feng Qianyi Chen Tailin Wu
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE102 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-19541

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

GenCP : Towards Generative Modeling Paradigm of Coupled physics Tianrun Gao Haoren Zheng Wenhao Deng Haodong Feng, Tao Zhang Ruiqi Feng Qianyi Chen Tailin Wu 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 “conditional-to-joint” sampling scheme. We evaluate our paradigm on a synthetic setting and three challenging multiphysics scenarios to demonstrate both principled insight and superior application performance of GenCP . Code is available at this repo: github.com/AI4Science-WestlakeU/GenCP . † † footnotetext: ∗ Equal contribution. † Work done as an intern at Westlake University. ‡ Corresponding author. 1 Introduction Most real-world physical systems are governed by the intricate interplay of multiple physical processes spanning diverse disciplines ( 51 ) . As a result, multiphysics problems are widely recognized as both fundamental and practically valuable. They arise across a broad range of applications, including aerospace engineering ( 33 ; 20 ) , biological engineering ( 38 ; 14 ) , and civil engineering ( 48 ; 50 ) . Despite their importance, accurate simulation of coupled physics remains notoriously difficult ( 34 ) , largely due to strong cross-physics interactions and the resulting high system complexity, which are far more challenging to model than in single-physics settings. Numerical multiphysics simulation methods include tightly coupled methods, which achieve high fidelity by solving the entire system jointly ( 24 ; 52 ) , often incur prohibitive complexity and computational costs in real-world applications ( 16 ) . Loosely coupled methods, which solve each physics field separately and iteratively exchange information until convergence, are more widely adopted, balancing practicality, efficiency, and acceptable precision ( 30 ) . However, both approaches suffer from requirements of extensive interdisciplinary expertise for solver development, and high computation demands ( 9 ) . Surrogate models ( 34 ) and neural operators ( 27 ) , therefore, have emerged as an alternative to accelerate multiphysics simulation, given the limitations of numerical methods. Most of these models, however, rely heavily on coupled solutions as training data, making data acquisition at least five times more computationally expensive compared to using decoupled data ( 7 ; 6 ) . To enable training on more accessible data from decoupled physics, surrogate models further borrow ideas from numerical solvers, such as the Gauss–Seidel framework ( 36 ) and perform ADMM-like ( 8 ) iterative inference to approximate the coupled solution ( 31 ; 13 ) . Although improving efficiency and leveraging decoupled data, surrogate models struggle with complex spatiotemporal dynamics due to their limited ability to capture high-frequency, high-dimensional, and stochastic behaviors ( 28 ) . The defect motivates emerging efforts in generative simulation ( 11 ) . Notably, existing generative approaches primarily focus on single-physics problems ( 43 ) or directly learn multiphysics solutions from coupled data ( 37 ) , largely overlooking the challenge of learning coupled physics from decoupled training data. A recent study ( 54 ) explored embedding coupling iterations into every denoising step of diffusion models to enable coupled sampling. However, the study lacks a rigorous theoretical foundation or guarantees of reliability. Emerging generative approaches show promise for high-fidelity modeling but typically rely on coupled data or lack theoretical guarantees of reliability. The above challenges raise a critical research question : How can we develop a framework that learns coupled physics from decoupled training data while ensuring high fidelity, efficiency, and reliability? To address the challenges, we introduce Gen erative C oupled P hysics Simulation ( GenCP ), a novel and principled paradigm for multiphysics simulation, which achieves decoupled training and coupled inference with “3H”, combining high fidelity from generative modeling, high efficiency from our “coupling in flow” design, and high reliability guaranteed by numerical theory. We begin by clarifying that the essence of solving coupled problems can be reformulated as modeling probability density evolution in functional spaces of physical fields. During training on decoupled data, we employ flow matching ( 29 ) to learn the evolution of conditional probability, where each physical field evolves from noise to its target conditioned on the other one. During inference, we apply operator splitting in the functional space at each flow step, thereby merging conditional probability transitions into a coherent joint distribution transition. Our method physically corresponds to iteratively solving the coupled fields in the noisy latent space as they evolve toward the solution. Figure 1 illustrates the overall schematic of the GenCP paradigm. Starting from the continuity equation of the joint distribution, we show that the flow matching ( 29 ) sampling process can be reinterpreted with an operator-splitting scheme for probability density functions. We prove in a Hilbert space that our proposed joint inference method enjoys error controllability guarantees. On top of the theoretical foundation, we examine GenCP on both naive and complex settings. First, we design a synthetic dataset in 2D space to demonstrate the joint distribution sampling effect from learning the conditional distribution. Then, we evaluate GenCP on three challenging multiphysics settings and compare it against surrogate-based and other generative paradigms. Across these tasks, GenCP consistently outperforms existing approaches with error reducing ranging from 12.54% to 42.85%, and even exceeding 65% on certain metrics, demonstrating its effectiveness and robustness in coupled scenarios. Figure 1: The schematic of GenCP . With a pretrained model with decoupled physics, GenCP could manage to achieve coupled physics during flow steps. Note that here we only use Lie-Trotter Splitting as a demonstration. The contributions of our work include: (1) We reformalize generative coupled simulation by modeling it as probability density evolution in the functional space of the physical field, offering a principled view of the problem. (2) We propose GenCP , an efficient generative paradigm that achieve coupled inference with decoupled learning by introducing an operator splitting mechanism into the flow matching process. (3) We establish a theoretical guarantee by connecting the flow-matching continuity equation with operator splitting, and prove that our scheme enjoys controllable error bounds. (4) We conduct concrete evaluations on both synthetic setting and challenging multiphysical scenarios, where GenCP consistently outperforms surrogate and generative baselines. 2 Related Work In this section, we briefly review the related work on both numerical and learning-based coupling simulation, highlighting key insights as well as the gaps that motivate our approach. A full discussion is provided in Appendix A . Numerical simulation. Classical numerical solvers are typically categorized into tightly coupled (monolithic) and loosely coupled (partitioned) schemes. Monolithic approaches enforce interface conditions exactly and ensure strong stability, but are computationally prohibitive ( 40 ; 44 ) . Partitioned approaches are more efficient and easier to implement, though they are often prone to unstable convergence (

Read the complete text