Propositions de projet de master: LMS

Deep geological repositories are the internationally preferred solution for the long-term disposal of high-level radioactive waste. Ensuring the integrity and safety of these systems over thousands of years requires robust monitoring of repository-induced effects—particularly temperature, pore water pressure, displacement, and stress—within the engineered and geological barriers.

This MSc thesis focuses on developing advanced, data-driven forecasting models using graph neural networks (GNNs) to predict the evolution of key parameters—temperature, pore water pressure, displacement, and stress—based on long-term, spatially distributed sensor data.

The available data is sourced from heterogeneous sensor modalities, including:

• Fibre optical cables (distributed temperature or strain sensing)
• Thermometers and humidity sensors
• Thermal conductivity probes
• Pressure cells
• Extensometers
• Gas sampling ports and gas sensors
• Corrosion coupons
• Seismic sources and receivers
• Geophysical monitoring pipes (GPx)

Key objectives include:

• Construct a graph-based representation of the repository nearfield, encoding spatial topology, sensor locations, and interdependencies.
• Integrate multi-modal time-series data into a unified framework for learning spatiotemporal correlations.
• Develop and benchmark GNN-based predictive models (e.g., spatiotemporal GCNs, graph attention networks, temporal GNNs) for forecasting THM evolution.
• Compare performance to classical time-series methods
• Apply uncertainty quantification techniques using ensemble learning or Bayesian GNNs to assess reliability.

This work bridges the gap between experimental monitoring and operational forecasting, providing a critical tool for ensuring the long-term safety and reliability of nuclear waste disposal systems.

Supervision: Prof. Lyesse Laloui [email protected] (LMS lab) & Prof. Olga Fink [email protected] (IMOS Lab)