How do you predict a material's properties when the number of possible atomic configurations exceeds 2^2500? That is the bottleneck our IRESNE (nuclear fuel physics) and LIST (AI) teams have just cracked with PULSE, a generative (VAE) method published in Nature Scientific Reports, already cutting computational cost by more than two orders of magnitude (22,282 CPU hours down to 85 on a test case). With no known equivalent in the international literature, PULSE positions CEA as a pioneer in generative sampling of the configuration space of chemically disordered materials.

This 24-month postdoc gives you the opportunity to drive this method toward its next generation, leading three ambitious, parallel research axes: pushing model accuracy on systems of several thousand atoms with an IWAE architecture; equipping it with the ability to quantify its own uncertainty — a prerequisite for any use in nuclear safety; and, in the second year, tackling a high-value exploratory axis — generalizing PULSE to a continuous latent space, opening the door to any disordered crystal or alloy.

You will work at the heart of an all-CEA consortium bringing together two complementary strengths — atomistic nuclear fuel physics at IRESNE and state-of-the-art generative AI at LIST — with access to CEA supercomputers, the freedom to publish in top-tier journals, and the prospect of seeing your results feed directly into reactor safety analyses through the PLEIADES platform. A position built for a curious mind who wants to combine cutting-edge generative AI research with concrete impact on a strategic nuclear-energy challenge.