Learning Where to Simulate: Generative Active Sampling for Online PDE Surrogate Training
Researchers have developed a new active learning method called Online Generative Active Sampling (OGAS) to improve the training of data-driven surrogate models for partial differential equations (PDEs). This method uses a diffusion model to learn and control the data sampling distribution, prioritizing configurations that lead to challenging dynamics for the surrogate. OGAS aims to reduce errors in the worst-case scenarios and improve overall error dispersion, with negligible overhead. AI
IMPACT Enhances the reliability of AI models used in complex scientific simulations, potentially accelerating discovery in fields relying on PDE solvers.