• atomistic modeling
• Behler Parrinello
• catalysis
• catalyst restructuring
• conformal sampling
• MACE
• machine learning
• methanol decomposition
• RGO
• sampling

In this work we explore machine learning accelerated modeling of potential energy surfaces of catalyst surfaces and their sampling. We compare different machine learning potentials, mainly the oldest one (Behler-Parrinello) and a state of the art equivariant graph convolutional one (MACE) assessing their accuracy in interpolating DFT and their computational cost. We build an ensemble of codes to run different sampling techniques (metadynamics, NEB, Pathsampling, Reactive Global Optimization) with the aforementioned potentials. We apply them to a Pt2Mn catalyst oxide induced restructuring process analyzing their potentialities, strengths and weaknesses. Finally we develop a protocol to conformally transfer the database of structures for a given reaction on a metal surface to a different metal/alloy (sampling in chemical space). We successfully test this procedure on the methanol decomposition reaction.