Huge thanks to the amazing collaborators Simone, Matthew, Luigi, Umberto, Andy —for pushing the boundaries of ML and electronic structure in catalysis.

Feedback and comments welcome!

3/ We applied this framework to C–H activation catalyzed by a transition metal carbide. Training MLPs on MC-PDFT data lets us simulate long-timescale catalytic dynamics with multireference accuracy—something not achievable with standard methods.

2/ Sampling is just as critical as accuracy. We combine WASP with enhanced sampling (OPES) and active learning (DEAL) to efficiently generate a minimal, diverse training set—capturing key regions like transition states while keeping data cost low.

1/ Multireference methods like MC-PDFT are key for strongly correlated systems, but training MLPs requires a consistent active space across diverse, uncorrelated geometries. WASP solves this using a weighted sum of MO coefficients to generate consistent initial guesses.