Data-Efficient Neural Operator Training via Physics-Based Active Learning
Researchers have developed a new active learning technique called physics-based acquisition to improve the efficiency of training neural operators for solving partial differential equations. This method uses the equation's residual to intelligently select the most informative data samples, reducing the overall data requirements for training. Experiments on the 1D Burgers equation and 2D compressible Navier-Stokes equations demonstrate that this approach outperforms random acquisition and matches state-of-the-art data efficiency while incorporating a physics-based inductive bias. AI
IMPACT Enhances data efficiency in training neural operators for scientific simulations, potentially accelerating discovery in fields relying on solving differential equations.