Physics-aware neural networks for hydrology
Embedding mass conservation into neural networks to study how physical constraints shape performance, generalization, and learned representations.
Purely data-driven models can be highly accurate but are not guaranteed to respect basic physical laws. This work develops physics-aware neural networks that embed mass-conservation principles directly into the model, and investigates how these physical constraints affect:
- Predictive performance
- Generalization across diverse hydroclimatic conditions
- Learned representations inside the network