Machine learning–driven advances in geophysical data assimilation
| When |
Sep 28, 2026
from 10:45 to 11:45 |
|---|---|
| Where | D8 001 Buisson |
| Contact Name | Antoine Venaille |
| Attendees |
Marc Bocquet |
| Add event to calendar |
vCal
iCal
|
Machine learning (ML), particularly deep learning (DL), is playing a growing role in geophysical data assimilation (DA), a cornerstone of numerical weather prediction that combines observations with model predictions to estimate the state of a system. ML can enhance existing methods, complement them, or potentially replace parts of the DA cycle.
This talk reviews recent developments and promising directions for integrating ML into DA to improve forecasting in the geosciences. ML can provide efficient approximations of complex physical processes, correct model errors, and help combine information from models and observations. In ensemble DA, which uses multiple forecasts to represent uncertainty, ML can improve how these forecasts are generated and how DA methods are tuned. It also opens the possibility of learning how to update forecasts with observations, or even learning the entire DA and forecast cycle.
I will illustrate these opportunities with two examples: using DL to discover new and efficient ways to combine forecasts with observations, and embedding generative AI within classical DA methods.