Learning Where to Simulate: Generative Active Sampling for Online PDE Surrogate Training

Researchers introduce Online Generative Active Sampling (OGAS), an active learning method that improves PDE surrogate models by strategically sampling challenging configurations during training. Using a parallel diffusion model to steer data generation toward difficult regimes, OGAS reduces worst-case prediction errors across multiple PDE types without significant computational overhead.

The core challenge addressed here involves training machine learning surrogates on partial differential equations—computationally intensive physics simulations used across engineering and scientific domains. Uniform sampling of training data typically underrepresents edge cases and challenging dynamics, resulting in surrogates that perform adequately on average but fail catastrophically in high-stakes scenarios. This reliability problem matters significantly for real-world deployment where worst-case performance often determines system safety.

OGAS represents an elegant solution to the exploration-exploitation trade-off in machine learning. By training a conditional diffusion model in parallel, the method learns implicit relationships between PDE parameters and surrogate difficulty, then actively samples from high-difficulty regions. This couples data generation with model training, enabling dynamic steering without offline preprocessing—a practical advantage for computationally expensive simulations.

The experimental validation across Kuramoto-Sivashinsky, Navier-Stokes, and Gray-Scott equations demonstrates the method's generality. Most importantly, OGAS achieves substantial improvements in tail statistics (99th percentile errors), with negligible wall-time overhead. The trade-off between average and worst-case performance is explicitly acknowledged rather than hidden, providing practitioners with informed decision-making criteria.

For scientific computing and engineering applications relying on surrogate models, this advance enables more robust deployments with minimal additional computational cost. The methodology could extend to other domains where model reliability across parameter distributions matters more than mean performance. Future work likely focuses on scaling to higher-dimensional PDE spaces and exploring alternative difficulty signals beyond loss and uncertainty.

• →OGAS uses a parallel diffusion model to actively sample challenging PDE configurations, improving worst-case surrogate performance by significant margins
• →The method reduces 99th percentile prediction errors while maintaining negligible computational overhead compared to uniform sampling approaches
• →Tested across multiple PDE types and up to 308 parameters, demonstrating robustness across distinct challenging dynamics
• →Acknowledges explicit trade-off between average error and tail reliability, enabling informed deployment decisions for practitioners
• →Online training coupling enables real-time steering of data generation without preprocessing delays in the surrogate workflow