Talk, Amphi Dorothy Hodgkin, ENS Paris-Saclay, Gif-sur-Yvette, France
This presentation represents my official PhD defense, addressing the critical challenge of predicting the lifetime of power electronic modules to enable effective predictive maintenance.
Poster, La cité, Toulouse, France
This poster was presented during the Mobilit.Ai forum to showcase the adaptation of several machine learning techniques such as active learning and dynamic time warping in the context of lifetime estimation of power electronic modules to develop predictive maintenance strategies.
Talk, Palais de la Bourse, Bordeaux, France
This presentation explores the use of machine learning-based metamodels to efficiently assess the reliability of power electronic modules, capitalizing on their fast inference speed and high predictive power. A major focus of the presentation is the thorough analysis of often-overlooked statistical properties that are crucial for evaluating model behavior, as well as presenting a novel simulation selection technique based on active learning.
Talk, Université Gustave Eiffel, Versailles, France
Presented to the SATIE team’s branch at Université Gustave Eiffel, this talk was specifically tailored for an audience of power electronics experts. The presentation introduced a machine learning-based perspective on the ongoing challenges of power module reliability. It showcased how data-driven tools, specifically stochastic methods and Dynamic Time Warping algorithms, can be effectively applied to model degradation and improve reliability assessments in this domain.
Poster, Conservatoire national des arts et métiers, Paris, France
This poster was presented during the SATIE lab’s annual doctoral students’ day to present my PhD’s work.
Talk, Arts et Métiers – ENSAM (Paris Campus), Paris, France
This talk took place at the 2025 Digital Twins in Engineering & Artificial Intelligence and Computational Methods in Applied Science confernece (DTE - AICOMAS 2025), it presented how machine learning can be used to create surrogate models for finite element simulations to reduce computational time. This work was carreid out to enable advanced frameworks for remaining useful life estimation.
Talk, ENS Paris-Saclay, Gif-sur-Yvette, France
This presentation was delivered to the STAN team at the LMPS laboratory, bridging the gap between applied machine learning and mechanical engineering. The talk introduced the mechanical community to data-driven approaches for a complex multi-physics problem: estimating the lifetime of power electronic modules. Specifically, it explored how machine learning can be leveraged to model bond wire degradation through a Markov chain-based damage modeling coupled with surrogate machine learning models to infer mechanical indicators based on numerical simulation data.
Poster, Paganini Conference Center, Parma, Italy
This poster was presented during the 35th European Symposium on Reliability of Electron Devices, Failure Physics and Analysis (ESREF 2024). Throught this poster, we show how kernel density estimation can be utilised as a probability density estimator, to be utilised in a Markov-Chain based sampling scheme. This sampling scheme allows us to capture the dynamic responsible for the power electronic modules’ failure, using experimental data of tests to failure, combined with numerical simulations used to generate mechanical features.
Talk, Amphi Dorothy Hodgkin, ENS Paris-Saclay, Gif-sur-Yvette, France
This talk was given during the SATIE lab’s annual doctoral students’ day to present my PhD’s work. This talk has recieved the award of the best presentation.
Talk, Four Points by Sheraton Catania Hotel & Conference Center, Aci Castello, Catania, Italy
This work was presented during the 2024 25th International Conference on Thermal, Mechanical and Multi-Physics Simulation and Experiments in Microelectronics and Microsystems (EuroSimE). Its main purpose is to highlight how a physics-based model (Paris’ law) and a data-driven model (adaptive polynomial regression) can be combined to obtain a hybrid model that benefits from both model types’ strenghts.