Machine learning (ML) is now commonly used in materials science to enhance the predictive capabilities of physical models. ML interatomic potentials (MLIP) trained on electronic-structure calculations have become standard tools for conducting efficient yet physically accurate molecular dynamics simulations. More recently, unsupervised generative ML models are being explored to learn hidden property distributions, and generate new atomic structures according to these distributions. This is useful for chemically disordered solid solutions, whose properties depend on the distribution of chemical species in the crystal lattice. In such cases, the number of possible configurations is so large that exhaustive sampling is beyond the capability of conventional methods. An example is U-Pu mixed oxides (MOX), a type of nuclear fuel that can significantly reduce the volume and radiotoxicity of spent-fuel waste. High-entropy alloys are another class of disordered materials that are promising candidates for radiation-resistant nuclear materials.

The goal is to combine MLIPs and generative methods to address atomic transport properties in MOX. The candidate will use in-house ML generative tools to generate representative atomic configurations and build an ab initio database. They will use this database to train a new MLIP for MOX, leveraging the experience gained from developing analogous MLIPs for the pure UO2 and PuO2 oxides. Finally, they will apply the MLIP to calculate atomic diffusion coefficients, crucial for predicting irradiation-induced microstructure evolution and in-reactor behavior.

The work will be conducted at the Nuclear Fuel Department (IRESNE, CEA Cadarache), within a scientific environment characterized by a high level of expertise in materials modelling, and in close collaboration with other CEA teams in the Paris region specialized in ML methods. Results will be disseminated through scientific publications and participation in international conferences