Machine learning and surrogate model for two-case studies in nuclear waste management

Innovations in machine learning (ML) provide new opportunities in the safety assessment and design optimization of deep geological repositories (DGRs) and waste containment systems. A study on the Full-scale Emplacement (FE) experiment at the Mont Terri Underground Laboratory demonstrates ML's potential to enhance 3D heat transport models. ML was applied to data, which were collected from sensors installed in the bentonite filled buffer zone between the heat producing waste canister and host rocks. A neural network-based surrogate model was developed, which accurately predicts thermal conductivity considering the evolution of local humidity, crucial for assessing temperature evolution in DGR tunnels. The Elman Neural Network outperforms other models, validating heat transport calculations against extensive temperature sensor data with minimal error, thereby improving DGR design and operation. Complementary, a neural network-based surrogate model, was applied to simulate geochemical interactions within cementitious waste packages over a period of 100 years. This approach significantly reduces computational time, offering insights into the deterioration mechanisms of materials like iron and aluminum. By evaluating 1 million evolution scenarios, a sensitivity analysis across this extensive dataset reveals that the hydrogen gas production is unique sensitive to various input variables, particularly corrosion rates, underscoring its importance in geochemical process evaluations. Together, these studies showcase ML's profound impact on advancing waste management technologies, ensuring environmental safety and operational reliability.