Ayat-Allah Bouramdane's research focuses on climate change, particularly its impacts, as well as mitigation and adaptation strategies. Her work relies on modeling, optimization, and decision-support tools to support energy, ecological, digital, and societal transition policies at both global and regional levels. An energy engineer by training, she completed exchange programs at the European School of Materials Engineering (EEIGM), the National Graduate School of Electricity and Mechanics (ENSEM) in Nancy, and the Faculty of Science and Technology (FST) in Nancy. She is a graduate of École Polytechnique and the Institut Polytechnique de Paris, where she conducted her doctoral research at the Laboratoire de Météorologie Dynamique – Pierre-Simon Laplace Institute (LMD-IPSL). She has teaching and research experience at several academic and research institutions.

How can AI-enabled digital twins predict fuel cell degradation while maintaining sovereign control over operational data across geographically distributed infrastructures?    How can artificial intelligence-powered digital twins predict fuel cell degradation while allowing each organization to keep its operational data under its own control, without sharing it with a central server?Fuel cells (devices that generate electricity by converting hydrogen and oxygen into energy, with water as the only by-product) gradually lose efficiency over time. For example, a fuel cell that produced 100% of its rated power when new may, after several years of operation, produce only 85% or 80% of that power while consuming more hydrogen to generate the same amount of electricity.However, it is difficult to predict exactly when this degradation will occur because every fuel cell operates under different conditions. For example, a fuel cell installed in a very hot and dusty region will not age in the same way as one operating in a cold and humid climate. Likewise, a fuel cell continuously supplying power to a hospital (24/7 operation) will experience different wear than one used only a few hours a day to power a charging station. The way the fuel cell is maintained (regular or irregular maintenance), variations in electricity demand (high or low loads), and the quality of the hydrogen used also influence its rate of degradation. This is why two identical fuel cells can have very different service lifetimes.To improve prediction accuracy, it is necessary to learn from data collected from a large number of fuel cells. However, organizations are often unable to share their data because it is confidential (it contains sensitive information), strategic (it provides a competitive advantage), or protected by regulations. This constraint is known as data sovereignty (the principle that an organization retains ownership, control, and responsibility for its own data).This research proposes a Sovereign AI Digital Twin (SAIDT). A digital twin is a virtual copy of a real fuel cell. It can be imagined as a living digital counterpart that continuously reproduces the condition and behavior of the fuel cell using information received from its sensors (temperature, pressure, voltage, electric current, hydrogen flow rate, humidity, etc.). For example, if the fuel cell temperature increases or its power output decreases, the digital twin immediately detects the change and updates its model.This digital twin uses artificial intelligence (AI), that is, algorithms capable of learning from data, to recognize the early signs of degradation. For example, if the AI detects that a combination of high temperature, voltage drop, and increased hydrogen consumption has previously led to failure in similar fuel cells, it can estimate that degradation is also occurring in the fuel cell it is monitoring and predict how this degradation will evolve.Unlike conventional approaches, the raw data (the measurements directly recorded by the sensors) is never sent to a central server. Instead, it remains stored within the organization operating the fuel cell. For example, a transportation company, a hospital, and a power plant each keep their own data on their own servers. This protects sensitive information related to their operations and activities.Instead, each digital twin shares only what it has learned, namely the parameters of the artificial intelligence model (the internal settings that enable the model to make predictions), through Federated Learning. This can be compared to a classroom where each student solves a different exercise at home: no student shows their own work to the others, but each explains the method they used. In the end, everyone benefits from the best methods without exchanging their individual exercises. Similarly, each organization helps improve the common model without revealing its confidential data.The system also relies on Edge Computing, meaning that computations are performed directly near the equipment, for example on a computer installed in a hydrogen station, inside a vehicle, or in a factory, rather than sending all the information to a remote data center. This approach provides faster results, reduces network communication, and allows the system to continue operating even when the Internet connection is slow or temporarily unavailable.In addition, the system incorporates physics-based models, that is, scientific equations describing the real behavior of a fuel cell (electrochemical reactions, gas transport, electricity generation, performance losses, etc.). For example, if the AI predicts a sudden performance drop while the laws of physics indicate that such a drop is impossible under the current operating conditions, the model can correct the prediction. Combining AI with physics-based models therefore produces more accurate and reliable predictions, even when only limited data is available.The main objective is to predict the Remaining Useful Life (RUL), that is, the estimated time before a fuel cell loses enough performance to require maintenance or replacement. For example, if the system estimates that a fuel cell can continue operating for approximately 1,500 hours before its efficiency becomes insufficient, the operator can schedule maintenance several weeks in advance, order the necessary replacement parts, and avoid an unexpected failure.The proposed framework will be tested using data collected from fuel cells operating in very different environments, such as hydrogen buses running in cities, power generation stations, data centers, and industrial facilities located in hot, cold, humid, or desert regions. This diversity will make it possible to verify that the system performs well under a wide range of real operating conditions.The ultimate goal is to develop a system that is secure (protecting data from unauthorized access), reliable (providing consistent and accurate predictions), explainable (able to explain why it predicts degradation, for example by showing that temperature, voltage drop, and hydrogen consumption are the main factors behind its diagnosis), and scalable (capable of being progressively deployed across hundreds or thousands of fuel cells distributed over different sites and countries).Such a system will make it possible to anticipate maintenance operations before failures occur, extend the lifetime of fuel cells, reduce operating costs, improve system availability, and accelerate the deployment of hydrogen technologies, while ensuring that every organization remains the full owner and controller of its own data.Keywords: AI Digital Twin (intelligent virtual replica), Fuel Cell Degradation (performance aging), Federated Learning (collaborative learning without data sharing), Data Sovereignty (local control of data), Predictive Maintenance (maintenance planned based on predictions).