NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·1h ago
AI Research
arXiv CS.AI·Jun 10·5 min read
Learning Where to Simulate: Generative Active Sampling for Online PDE Surrogate Training
TL;DR
Generating training data for surrogate models of partial differential equations (PDEs) is challenging due to under-representation of complex dynamics. Online Generative Active Sampling (OGAS) was developed to improve data sampling by focusing on difficult configurations.
✦ Why It Matters
Engineers can implement OGAS to improve the reliability of surrogate models in complex simulations with minimal overhead.
Key Takeaways
- Use OGAS to enhance surrogate training for PDEs.
- Focus on challenging configurations to reduce extreme prediction errors.
- Leverage real-time data generation adjustments for efficiency.
- Evaluate performance improvements across various PDE scenarios.
How It Works
OGAS operates by training a diffusion model alongside the surrogate, which learns to map a difficulty signal—such as loss or uncertainty—to configuration parameters. This allows OGAS to actively sample data from regions of high difficulty, ensuring that the training set is representative of challenging dynamics.