KIMoDA: Opportunities and Limitations of AI-Driven Simulation Methods for Supporting Long-Term Safety Assessment of Deep Geological Disposal

The KIMoDA (Artificial Intelligence for Modelling Diffusive/Advective Flow in Porous Media) project aims to explore the opportunities and limitations of applying AI-driven simulation methods for the methodological development and evaluation of modelling approaches relevant to long-term safety assessments of deep geological disposal of high-level waste in Germany. The project investigates transient advection and diffusion processes in porous and fractured media, across kilometer-scale domains and timeframes of up to 1,000,000 years, where the use of conventional numerical solvers is computationally challenging.

On the technical side, KIMoDA evaluates AI-based surrogates and hybrid models, including deep learning architectures and physics-informed neural networks (PINNs), and, for the same calculation case, compares the results of these models against the results obtained using PFLOTRAN, an open-source solute transport simulation code. To this end, standardised reference cases are developed for the three host rock types considered in the German site selection process: claystone, rock salt and crystalline rock. Performance is assessed using metrics such as RMSE, R², and maximum norm error.

Interwoven with the development and benchmarking of AI surrogates and hybrid models, KIMoDA assesses socio-technical and ethical risks and opportunities that may arise when different AI systems are considered or discussed by experts and officials embedded in organisational, societal, and political contexts. Taking key requirements of the German Site Selection Act (StandAG) as a reference framework—traceability, reproducibility, accountability and participation, and precaution under deep uncertainty – the analysis examines how model properties (e.g., data dependence, opacity, bias, non-determinism) may interact with data governance, validation practices and use-patterns to shape trustworthiness. Explainable AI, sensitivity analyses, and targeted visualisations are applied to strengthen auditability and communicability of results for safety-critical decision-making. By uniquely combining AI modelling with socio-technical and ethical perspectives, KIMoDA aims to contribute to the development of methodological approaches relevant to reproducible, explainable, and publicly credible AI-supported safety cases in nuclear waste management.