PI-JEPA: Closing the Data Asymmetry in Subsurface Surrogate Modeling via Label-Free Pretraining

Brandon Yee, Pairie Koh, [4pt] Physics Lab, Yee Collins Research Group · 2026

Read PI-JEPA: Closing the Data Asymmetry in Subsurface Surrogate Modeling via Label-Free Pretraining by Brandon Yee, Pairie Koh, [4pt] Physics Lab, Yee Collins Research Group (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-2604-01349.txt.

Description

Yee and Koh introduce PI-JEPA, a pretraining framework for surrogates of multiphysics simulations, learning without a single completed PDE solve. It trains by masked latent prediction on generated unlabeled parameter fields, under per-sub-operator PDE residuals. Physics-constrained modules handle pressure, saturation transport and reaction, aligned to a Lie-Trotter split, so fine-tuning needs as few as 100 labeled runs. On single-phase Darcy flow it cuts error by 1.9 times against FNO and 2.4 times against DeepONet at 100 labeled runs, improving 24 percent over supervised-only training at 500 runs.

Catalog record

CREATORBrandon Yee, Pairie Koh, [4pt] Physics Lab, Yee Collins Research Group
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE85160
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.01349

Browse the full library · About the Tower of Babel