The challenge of multiphysics systems simulation that incorporates heterogeneous physical models with distinct spatial, temporal, and computational traits is the lack of high fidelity simulation, scalability, and interpretability. Relying on rules, domain-specific coupling strategies, or a crafted approach, traditional fusion techniques stand little to no chance of sustaining adaptability across multiple problem domains. To address this gap, we propose a learning framework called MetaFusion-X. This framework can learn optimal integration strategies to disparate physics models and thus solve them on a system-level understanding, facilitating efficiency and adaptability. Unified under deep neural operators and attention cross-domain mechanisms, the meta-learner architecture of MetaFusion-X forms a meta-learner framework. This operates in two phases: initially, wielding generalizable fusion priors through meta- training on a diverse scope of synthetic and real-world multiphysics scenarios, and later honing a specific task integration policy incurring minimal adaptation cost. Enabling capture of Lagrange relations, this paradigm allows dynamic shifting of latent representations of different processes and ease of interaction across scales and domains. Along with modulating fusion layers and adaptation heads for increased accuracy, these components of MetaMicroX aid with a hybrid loss function to maintain stability, interpretability, accuracy, and, most prominently, ease of use. MetaFusion-X leverages the ability to retarget pre-trained models on distinct tasks of cross-domain self-similarity, enabling seamless interactions across scales and domains. Empirical results demonstrate that MetaFusion-X outperforms predictive fusion models in terms of accuracy, computational efficiency, and broad applicability to diverse domains, like thermo-fluid interaction, electro-mechanical systems, and geophysical simulations. In addition, how meta-learned integration strategies are interpretable exposes hidden relationships and dependencies that are not visible with traditional approaches. These advancements mark another step towards integrated multiphysics reasoning. At the same time, scientific computing, digital twin engineering, and intelligent simulation design stand to benefit from MetaFusion-X’s promise as a core enabling technology.
Sreenivasu et al. (Mon,) studied this question.