Bridging Structure and Activity in Nanocatalysts via Machine Learning and Global Structure Representations

Establishing a mapping between nanocatalysts structure and their catalytic properties is essential for efficient design. To this end, we develop a general machine learning framework, validated on a representative and challenging application: predicting the mass activity of Pt nanoparticles for the electrochemical oxygen reduction reaction, estimated via a microkinetic model. Accurate models are obtained when leveraging either a nanocatalyst's structure representation accessible at the computational level, namely the surface site generalized coordination number distributions, or one accessible experimentally, namely the nanoparticle's pair distance distribution function. The same representations also enable to predict with a good accuracy the stability of the same nanoparticles. Building on this result, we demonstrate that our machine learning approach, in tandem with multi-objective Bayesian optimization, efficiently identifies with a large pool, candidate nanocatalysts that display an optimal trade-off between energetic stability and activity. These findings provide a robust blueprint for accelerated theoretical and experimental identification of active nanocatalysts.