KAN4Met: Interpretable Networks for Weather Forecasting

Neural networks are good at forecasting and bad at explaining themselves. Kolmogorov–Arnold Networks (KANs) are a newer kind of network where the learned functions sit on the connections, so you can plot them and read what the model learned.

KAN4Met is a research project on KANs for weather forecasting. It is funded by Spain’s Agencia Estatal de Investigación (PID2024-155187OB-I00), led by David Gómez-Ullate, and runs from July 2025. I’m part of the team. I’m interested in how they can help weather routing for ships.

This is the youngest of my research lines, so there are no results to show yet. The plan is to connect it with my other work: fast, readable stand-ins for expensive ship models in the routing, and exposure-response curves learned from data in the climate-risk models. Read the news post.