Séminaire des doctorant·es SO

par Dr Alexey Lazarev (IMT, ANITI), Daphné Matoses (UT, IMT, INSERM, IRESP), Margot Ferré (ENAC, IMT), Naomi Albukerque (IMT, Laplace), Younes Essafouri (IMT, CNRS)

mardi 16 juin 2026, 11:15 → 12:15 Europe/Paris

Salle K. Johnson (1R3, 1er étage)

Alexey Lazarev (IMT, ANITI)

Geometry of latent spaces: curvature regularization and Riemannian clustering

Two geometric contributions for autoencoder latent spaces are presented. The first is a curvature regularization framework that augments the autoencoder loss with curvature-based penalties and auxiliary regularizers to promote structured and interpretable latent representations. The second is a Riemannian k-means algorithm that leverages geodesic approximations derived from a Faber–Schauder basis. Together, these approaches incorporate geometric information into both representation learning and clustering, enabling a deeper understanding and exploitation of latent-space structure.

---

Naomi Albukerque (IMT, Laplace)

Physics-Informed Learning for Charge Transport Characterization in Dielectric Materials

Electrical transport and space charge accumulation in insulating materials play a key role in the performance and reliability of high-voltage electrical systems. Characterizing the transport mechanisms within a material from experimental observations remains a challenging inverse problem. This work addresses the inverse problem of identifying the parameters of a PDE-based charge transport model from experimental measurements. The proposed approach combines regularized deconvolution techniques and Physics-Informed Neural Networks (PINNs) to exploit both measurement data and prior physical knowledge. Charge density profiles are first reconstructed from indirect observations obtained through the Pulsed Electro-Acoustic (PEA) method. These profiles are then used to train a PINN that embeds the governing partial differential equations into the learning process, enabling the simultaneous estimation of state variables and unknown model parameters. Finally, a global sensitivity analysis will be performed to quantify the influence of the identified parameters on the model outputs.

---

Daphné Matoses (UT, IMT, INSERM, IRESP)

Causal inference and synthetic patients in randomized trials

Randomized Controlled Trials (RCTs) are the gold standard for causal inference, but they suffer from small sample sizes, high costs, and ethical constraints. To estimate the Average Treatment Effect (ATE), several estimators are available, from simple difference-in-means to doubly robust methods such as AIPW, all relying on standard identification assumptions. A natural idea is to supplement the trial with synthetic data to improve these estimations. We show that endogenous synthetic data, generated from the trial itself, bring no additional Fisher information about the ATE. Exogenous synthetic data can help, but only if the generative model is accurate enough, which raises the question of what makes a good generator.

---

Margot Ferré (ENAC, IMT)

Stochastic Mirror Descent & Application to the estimation of Sobol' indices

Sobol' indices are widely used in Global Sensitivity Analysis to quantify the sensitivity of a random model with respect to its input variables. Classical estimators, such as the Pick-Freeze estimator, are asymptotically efficient but estimate only one index at a time. This poster presents how Stochastic Mirror Descent (SMD) algorithms enable the simultaneous estimation of all Sobol' indices. We establish general convergence results for SMD under strong convexity assumptions, and discuss the work of Gadat, Costa, Gendre and Klein (2025), which formulates Sobol' index estimation as a constrained optimization problem solved through a constrained SMD algorithm. Based
on negative entropy, this scheme yields an explicit estimator with almost sure convergence guarantees. Future work focuses on deriving Central Limit Theorems for general SMD estimators.

---

Younes Essafouri (IMT, CNRS)

A Framework for Explainable AI in Weather Forecasting: Diagnosing Deep Learning Models via Gradient-Based Attributions

Each day, potentially critical decisions made by governments and organizations depend on accurate weather forecasts, from deciding whether to evacuate ahead of a storm to simply carrying an umbrella. In this context, Deep Learning (DL) models are emerging as a popular and  computationally efficient alternative to traditional Numerical Weather Prediction (NWP) models, offering the potential to capture complex data patterns that explicit physical equations may miss (lam2023). However, their opaque (black-box) nature remains a barrier to operational trust. Explainable AI (XAI) seeks to address this opacity by revealing the decision process behind predictions. Indeed, classical XAI techniques can expose when DL models rely on spurious corrélations rather than causal physical mechanisms (geirhos2020). Yet their direct application to meteorological data often produces attribution maps that are noisy (kim2019) and difficult to interpret due to their high dimensionality. It also remains unclear whether these tools can consistently identify the complex physical drivers inherent in NWP (bommer2024).

Building on previous work (bommer2024, kim2023, yang2024), we establish a framework to generate compact and interpretable explanations of local weather forecast predictions produced by deep neural networks. These explanations build on the output of gradient-based methods such as VanillaGrad and SmoothGrad (smilkov2017), which scale well to high-dimensional data. More specifically, our framework enables targeted analysis by selecting a region of interest (e.g., the Paris area) and a target variable (e.g., accumulated precipitation). It therefore answers the question: “Why did the neural network predict this feature at this location?” To do so, it first computes dense attribution maps with respect to all input variables (e.g., wind components at varying altitudes). Traditionally, bounding boxes are used to define the region of importance in these maps (kim2023), but they cannot convey detailed directional information. We therefore propose determining régions of importance using “confidence ellipses” that summarize the center, principal directions, and importance of the most concentrated regions. Unlike bounding boxes, these ellipses displayed over the raw attribution maps as a background, provide rich and easily interpretable information about the directionality and spatial spread of the model’s focus.

Preliminary results on the hybrid transformer-convolutional model UNETR++ (shaker2024), trained and tested on the TITAN dataset from Météo-France (comprising hourly surface and vertical profiles of wind, temperature, and geopotential over metropolitan France), demonstrate our framework’s relevance for explaining predictions from deep neural networks. We were able to verify that different trained models successfully capture the vertical hierarchy of atmospheric variables, evidenced by an effective receptive field that expands with increasing altitude. More interestingly, our framework allowed us to identify systematic biases learned during training that correlate with known physical phenomena. These findings serve as a foundational step toward future work on developing novel explainability methods to detect whether trained models capture complex physical mechanisms.