PI-JEPA performs label-free pretraining of operator-split latent predictors on unlabeled parameter fields with per-sub-operator PDE residuals, enabling fine-tuning of multiphysics surrogates with 100 labeled runs and 1.9-2.4x error reduction versus FNO and DeepONet.
By Theorem 2.1 of Wainwright [2019], the minimax rate for estimating ann×n matrix fromN ℓ noisy linear measurements inR n isE[∥ ˆA−A∥ 2 F ]≥c·n 2σ2/Nℓ for a universal constantc >0
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
PI-JEPA: Label-Free Surrogate Pretraining for Coupled Multiphysics Simulation via Operator-Split Latent Prediction
PI-JEPA performs label-free pretraining of operator-split latent predictors on unlabeled parameter fields with per-sub-operator PDE residuals, enabling fine-tuning of multiphysics surrogates with 100 labeled runs and 1.9-2.4x error reduction versus FNO and DeepONet.