Talk · 2026
Explainable AI for Data-Driven Weather Forecasting
Doctoral seminar of the Stat&Opt team, Institut de Mathématiques de Toulouse
Publications
-
NeurIPS 2026
Explanation of Dynamic Physical Field Predictions using WassersteinGrad: Application to Autoregressive Weather Forecasting
Perturbing the input of a neural weather model does not make its gradient explanation noisier; it makes it move. We show that stochastic input perturbations geometrically displace attribution maps on dynamic physical fields, an effect that compounds under autoregressive rollout, and replace SmoothGrad’s pointwise averaging with an entropic Wasserstein barycenter that recovers a spatially coherent geometric consensus.
@inproceedings{essafouri2026wassersteingrad, title = {Explanation of Dynamic Physical Field Predictions using WassersteinGrad: Application to Autoregressive Weather Forecasting}, author = {Essafouri, Younes and Raynaud, Laure and Drozda, Luciano and Risser, Laurent}, booktitle = {Advances in Neural Information Processing Systems}, year = {2026}, eprint = {2604.22580}, archivePrefix = {arXiv} } -
EGU 2026 · Poster
A Framework for Explainable AI in Weather Forecasting: Diagnosing Deep Learning Models via Gradient-Based Attributions
@inproceedings{essafouri2026framework, title = {A Framework for Explainable AI in Weather Forecasting: Diagnosing Deep Learning Models via Gradient-Based Attributions}, author = {Essafouri, Younes and Seznec, Corentin and Drozda, Luciano and Raynaud, Laure and Risser, Laurent}, booktitle = {EGU General Assembly 2026}, address = {Vienna, Austria}, note = {EGU26-4039}, doi = {10.5194/egusphere-egu26-4039}, year = {2026} }
Talks & Teaching
Teaching · 2025–26
Linear models and statistical inference
Practicals for students at Université Paul Sabatier, from least squares in R to confidence intervals.
About
Neural networks now forecast the weather about as well as the physics-based models they learned from. I try to understand how they use physical information: which parts of the atmosphere a forecast relies on, how that influence travels from one forecast step to the next, and whether it makes physical sense.
PhD in explainable AI for weather forecasting at the Institut de Mathématiques de Toulouse, since 2025, with researchers at Météo-France and Cerfacs and support from ANITI. Before that, applied mathematics and scientific machine learning.
Path
-
2022 – 2025
Grenoble INP – Ensimag
Studied applied mathematics and computer science, from numerical solvers for PDEs to statistical learning.
-
Summer 2024
Laboratoire Hubert Curien
Replaced an expensive solver for the Kuramoto–Sivashinsky equation with physics-informed networks and Fourier neural operators.
-
2024 – 2025
KTH, Stockholm
Studied machine learning and worked on physics-informed models for fluid dynamics with the KTH AI Society.
-
2025
Dassault Systèmes
Developed physics-informed neural network surrogates for cardiac heat-transfer simulations.
-
2025 – now
PhD · Institut de Mathématiques de Toulouse
Opening up neural weather models: which parts of the atmosphere a forecast relies on, and whether that makes physical sense.
Off the clock I take photographs, read the novel before the film, and play records that never left the twentieth century.
- SteinbeckJohn Steinbeck, United States, 1902 to 1968
- García MárquezGabriel García Márquez, Colombia, 1927 to 2014
- HesseHermann Hesse, Germany and Switzerland, 1877 to 1962
- DazaiOsamu Dazai, Japan, 1909 to 1948
On rotation