Mechanism- and Data-Driven Exploration of a Global Descriptor for CO₂ Reduction

Atomic-scale structure-performance relations offer fundamental principles for catalyst design and optimization, where the descriptor plays a determining role. However, currently developed descriptors mainly focus on local information that fails in many typical and important cases, leaving huge gaps between experiments and computations. Herein, we successfully constructed a global descriptor to unveil the size effect of Cu nanoparticles (NPs) on the catalytic performance for CO₂ reduction reaction (CO₂RR), using a mechanism- and data-driven approach. Mechanism analysis suggests surface oxidation as a key global property to correlate the microscopic structure and macroscopic performance of NPs. A multiscale neural network framework, namely, ScaleNet, was proposed to realize the prediction of *OH coverage over NPs with experimental scale-size that cannot be processed by density functional theory (DFT). The integration of global and local information extractors helps ScaleNet accurately understand the surface adsorption behavior of NPs at different coverage levels, endowing this framework with excellent accuracy and extrapolation ability. Using this framework, *OH coverage over a series of Cu NPs with experimental scale-size were predicted, exhibiting strong correlation with the experimentally observed activity and selectivity. This supports the reliability of *OH coverage as a global descriptor, providing valuable insights and a novel learning paradigm for future explorations in nanoscale research.