Physics-informed Gaussian processes (GPs) offer a principled way to perform scientific inference while expressing uncertainty over measurements, boundary conditions, and governing equations. But as the abstract of a new arXiv paper notes, they struggle to scale. The paper, titled "Learn Only Where You Can't Compute: Structured Corrections to Gaussian Process Posteriors," appears to address this limitation directly.

The title hints at a strategy: instead of applying corrections everywhere, the method learns them only where direct computation is infeasible. This could reduce the computational burden while preserving the benefits of physics-informed GPs. However, the abstract is truncated, so specific algorithmic details are not available from the source.

Given the single source, we can only report what the abstract states: the framework exists, the scaling problem is acknowledged, and the proposed solution involves structured corrections. Readers interested in the full method will need to consult the paper directly.