Machine learning (ML) is nowadays of common use in materials science to improve the predictive capabilities of physical models. ML interatomic potentials (MLIP) trained on electronic-structure calculations have become standard tools for running efficient yet physically accurate molecular dynamics simulations. More recently, generative deep-learning models are explored to learn a hidden property distribution in an unsupervised manner, and generate new structures according to this hidden distribution. This is of particular interest for chemically disordered solid solutions, whose properties depend on the distribution of chemical species in the crystal lattice. In this case, the number of possible configurations is so high that an exhaustive sampling of this space is beyond the reach of conventional methods. An example is given by uranium-plutonium mixed oxides (MOX), a type of nuclear fuel that allows for a drastic reduction of the volume and radiotoxicity of spent-fuel waste.
The goal of this project is to combine MLIPs and generative methods to address atomic diffusion properties in MOX. The candidate will explore the use of various models (e.g., autoencoders) to find the collective variables that best describe the atomic diffusion trajectories, and to select the most representative atomic structures among the many possible ones. The most appropriate model will then be used to determine the diffusion coefficients of host species and fission products, which are essential for predicting the irradiation-induced microstructure evolution and thus the in-reactor behaviour of MOX fuels.
The work will be carried out at the Nuclear Fuel Department of the CEA, in a scientific environment characterised by a high level of expertise in materials modelling, and in close collaboration with other CEA teams in the Paris region specialised in machine-learning methods. The results will be promoted through scientific publications and participation in international conferences.