Active Learning of Moment Tensor Potentials for Atomistic Simulation of Pb Systems: A Small-Cell Active-Learning Framework Toward Fe-Pb Liquid-Metal Corrosion Modeling

Sammanfattning

Interatomic potentials provide an efficient way to model atomic interactions in large-scale simulations, but their predictive value depends on how well they reproduce the underlying potential energy surface across relevant structures and thermodynamic conditions. For liquid lead, a reliable potential is particularly important because accurate atomistic simulations are needed to study transport, phase behavior, and corrosion-related phenomena relevant to lead-cooled fast reactor applications. This thesis develops an automated small-cell active-learning workflow for constructing a Moment Tensor Potential for pure Pb. The model is trained on an initial dataset, used to explore configuration space through molecular dynamics, and iteratively refined by adding configurations identified as extrapolative through the active-learning criterion. The active-learning results show consistent convergence behavior: the training set grows from iteration to iteration, the number of newly selected extrapolative configurations decreases, and the root-mean-square errors for energies and forces generally improve. The validation tests further show excellent agreement for the melting temperature prediction, while the equation of state and elastic constants reveal more noticeable deviations from reference data, indicating that these properties remain more challenging to capture with the present training protocol. Overall, the work demonstrates a reproducible and extensible route for developing machine-learning interatomic potentials for Pb with minimal manual intervention. The resulting workflow is well suited for future simulations of lead-rich systems, including studies motivated by lead-cooled fast reactor technology.