PhD defense of Edgar JABER

Edgar JABER

Hybrid prognostics using simulation codes and statistical models applied to the study of steam generators clogging.

Abstract

This PhD thesis focuses on the development of hybrid methods for degradation prognostics in industrial systems. The main application concerns the clogging of steam generators (SGs) in pressurized water reactors operated by Électricité de France (EDF).

Two main families of models are used in prognostics: physics-based models and statistical, data-driven models. Hybrid approaches aim to combine both to leverage their respective strengths and improve robustness, particularly regarding predictive uncertainty. These tools are essential for planning maintenance or deciding on component replacement in long-lifespan critical infrastructure.
In scenarios where degradation data is regularly available over time, filtering techniques (e.g., Kalman, particle filters) are effective at correcting simulation-based predictions using new observations. However, for complex systems like SGs, sparse data and complex physics make the prediction task strongly context-dependent. Still, general methodological principles can be established.
Physics-based models come with structural biases and parametric uncertainty due to incomplete knowledge of input variables. Their use requires sensitivity analysis and rigorous uncertainty quantification (UQ), assuming the physical process is well modeled. When simulations are computationally expensive, surrogate models (or emulators) become necessary. The first part of this thesis develops a non-intrusive UQ method applied to an industrial clogging prediction code developed by EDF. Results align with expert knowledge and reveal significant prognostic uncertainty.

It is then crucial to evaluate the predictive quality of the emulators. Conformal prediction offers a robust distribution-free framework to construct prediction intervals with guaranteed coverage. We develop estimators suited for limited-data settings, producing intervals for scalar Gaussian processes. Unlike Bayesian credible intervals, our bounds are less sensitive to prior misspecification. For deterministic codes, interval width reflects surrogate approximation error, making them useful diagnostic tools.

The next phase involves conditioning the prior distributions on available heterogeneous data to improve predictive robustness. Unlike standard Bayesian calibration, typically applied using lab data, the goal here is to adapt probabilistic predictions to operational field contexts. We propose a data fusion approach inspired by data assimilation, tailored to sparse and heterogeneous sources (e.g., operator measurements, statistical models). Applied to a synthetic crack propagation case and SG clogging, the method significantly improves predictive performance. Open questions remain regarding latent variable uncertainty and discrepancy modeling.

Finally, in an exploratory direction, recalibrated simulations can generate degradation trajectories suitable for time series learning. Real-world systems often rely on sensor data that do not directly measure degradation. A key research question is whether such exogenous signals can predict future degradation states. If correlation exists, unobserved degradation may be inferred through features extracted from sensor signals.

This work contributes to the development of digital twins, where hybrid modeling, uncertainty quantification, and data integration enable the construction of robust and certifiable predictive frameworks for industrial components such as those in nuclear power plants.

Keywords

Prognostics, Hybrid models, Scientific computing, Uncertainty quantification, Machine learning,

Supervision

• Mathilde MOUGEOT
• Didier LUCOR
• Vincent Chabridon

Jury

• Eleni CHATZI, Full professor, ETH Zürich, Rapporteur
• Josselin GARNIER, Professeur des universités, Ecole Polytechnique, Examinateur
• Béatrice LAURENT-BONNEAU, Professeur des universités, INSA Toulouse, Rapporteur
• Pietro Marco CONGEDO, Directeur de recherche, INRIA, Examinateur
• Emmanuel VAZQUEZ, Professeur, Université Paris-Saclay, CNRS, CentraleSupélec, Examinateur

Publications