Scientific Machine Learning SciML ★
Definition
The umbrella term for the intersection of machine learning with scientific computing and physical modeling: both using AI to accelerate or replace simulation, and embedding physical knowledge into model structure. The organizing concept behind everything indexed on this site.
What it is
SciML is not one technique but a set of positions: physical law should enter the model (as a hard or soft constraint), simulation data should be fully exploited, and scientific evaluation criteria should replace pure ML accuracy chasing. PINNs, neural operators, reduced-order models and differentiable physics — all indexed here — live under this umbrella.
Key points
- Two threads: (a) AI for Science — use AI to accelerate discovery and simulation; (b) Science for AI — use physical structure to improve sample efficiency and generalization.
- How it differs from generic ML: SciML values sample efficiency, physical consistency, conservation and interpretability over raw task accuracy.
- Three gates to production: verifiable accuracy, quantifiable error, and a workable fallback. All three are required.
- Terminology: SciML covers Physics AI, which covers PINNs. In practice "Physics AI" leans engineering while "SciML" leans academic.
Common pitfalls / boundaries
- "Using a neural network to solve a PDE" is not automatically SciML; whether physical constraints are included is the dividing line.
- The literature grows extremely fast — prefer results with public code and benchmark numbers.