Prediction models that respect physics
Could a model make physically sound predictions from little data?
Where measurements are hard and data is scarce, we think a model trained on data alone can easily give physically wrong answers outside the range it has seen. Physics-informed neural networks (PINNs) were introduced in a 2017 paper to build physical laws into training. We want to explore that direction and share what we try.
Governing equations and observations
We aim to connect and clean equations and data, then have agents train models and verify each other.
A fast surrogate model
| Field | Physics |
|---|---|
| Status | Planned |
| Input | Governing equations · observations |
| Output | Fast surrogate model |
| Delivery | Joint research |