PI-JEPA: Label-Free Surrogate Pretraining for Coupled Multiphysics Simulation via Operator-Split Latent Prediction
arXiv:2604.01349v1 Announce Type: new Abstract: Reservoir simulation workflows face a fundamental data asymmetry: input parameter fields (geostatistical permeability realizations, porosity distributions) are free to generate in arbitrary quantities, yet existing neural operator surrogates require large corpora of expensive labeled simulation trajectories and cannot exploit this unlabeled structure. We introduce \textbf{PI-JEPA} (Physics-Informed Joint Embedding Predictive Architecture), a surrogate pretraining framework that trains \emph{without any completed PDE solves}, using masked latent prediction on unlabeled parameter fields under per-sub-operator PDE residual regularization. The predictor bank is structurally aligned with the Lie--Trotter operator-splitting decomposition of the governing equations, dedicating a separate physics-constrained latent module to each sub-process (pressure, saturation transport, reaction), enabling fine-tuning with as few as 100 labeled simulation ru
arXiv:2604.01349v1 Announce Type: new Abstract: Reservoir simulation workflows face a fundamental data asymmetry: input parameter fields (geostatistical permeability realizations, porosity distributions) are free to generate in arbitrary quantities, yet existing neural operator surrogates require large corpora of expensive labeled simulation trajectories and cannot exploit this unlabeled structure. We introduce \textbf{PI-JEPA} (Physics-Informed Joint Embedding Predictive Architecture), a surrogate pretraining framework that trains \emph{without any completed PDE solves}, using masked latent prediction on unlabeled parameter fields under per-sub-operator PDE residual regularization. The predictor bank is structurally aligned with the Lie--Trotter operator-splitting decomposition of the governing equations, dedicating a separate physics-constrained latent module to each sub-process (pressure, saturation transport, reaction), enabling fine-tuning with as few as 100 labeled simulation runs. On single-phase Darcy flow, PI-JEPA achieves $1.9\times$ lower error than FNO and $2.4\times$ lower error than DeepONet at $N_\ell{=}100$, with 24\% improvement over supervised-only training at $N_\ell{=}500$, demonstrating that label-free surrogate pretraining substantially reduces the simulation budget required for multiphysics surrogate deployment.
Executive Summary
This article introduces PI-JEPA, a novel surrogate pretraining framework for coupled multiphysics simulation. PI-JEPA leverages unlabeled parameter fields and operator-splitting decomposition to train a physics-constrained latent module for each sub-process. The framework achieves significant improvements in accuracy and reduces the simulation budget required for deployment. PI-JEPA outperforms existing methods, such as FNO and DeepONet, on single-phase Darcy flow simulations, demonstrating its potential in reservoir simulation. The authors' innovative approach to label-free surrogate pretraining has far-reaching implications for the field of multiphysics simulation.
Key Points
- ▸ PI-JEPA introduces a novel framework for label-free surrogate pretraining in multiphysics simulation.
- ▸ The framework leverages unlabeled parameter fields and operator-splitting decomposition for improved accuracy.
- ▸ PI-JEPA achieves significant improvements in accuracy and reduces the simulation budget required for deployment.
Merits
Strength
PI-JEPA's innovative approach to label-free surrogate pretraining enables the exploitation of unlabeled parameter fields, reducing the simulation budget required for deployment.
Demerits
Limitation
The framework requires a detailed understanding of the Lie--Trotter operator-splitting decomposition, which may limit its applicability to users without a strong background in multiphysics simulation.
Expert Commentary
PI-JEPA represents a significant advancement in the field of multiphysics simulation, offering a novel solution to the challenge of label-free surrogate pretraining. By leveraging unlabeled parameter fields and operator-splitting decomposition, the framework achieves improved accuracy and reduces the simulation budget required for deployment. While the framework requires a strong background in multiphysics simulation, its innovative approach has far-reaching implications for the field. As the demand for accurate and efficient multiphysics simulation continues to grow, PI-JEPA is poised to make a significant impact on industries and research communities relying on these methods.
Recommendations
- ✓ Further research is needed to explore the applicability of PI-JEPA to more complex multiphysics simulations and to develop user-friendly interfaces for non-experts.
- ✓ The development of label-free surrogate pretraining frameworks like PI-JEPA should be encouraged through funding and research initiatives, enabling the continued advancement of multiphysics simulation.
Sources
Original: arXiv - cs.LG
