Machine-learning-accelerated discovery of a σ-channel C-O activation mechanism

Carbon dioxide (CO₂) electroreduction is a promising route for carbon recycling and energy regeneration, with the development of dual-atom catalysts (DACs) being critical to its advancement. Conventional density functional theory (DFT) calculations and machine learning (ML) models based on DFT-derived descriptors are plagued by prohibitive computational time and resource costs. Herein, we extract 20 fundamental chemical descriptors exclusively from structural data and deploy a lightweight ML model to a dataset of 250 880 entries, eliminating the need for 12 544 DFT computations and rapidly identifying 10 high-performance two-dimensional BN-based DACs for CO₂ electroreduction. The model's reliability and the catalyst's stability and performance are validated via DFT, ab initio molecular dynamics (AIMD) simulations and literature evidence. Contrary to the classic π-backbonding mechanism that relies on strong CO adsorption to trigger activation, we reveal an unprecedented σ-donation-enabled pathway that achieves efficient CO activation via reduced C-O σ bond order within weak-adsorption regions. This defining σ-channel strategy decouples the intrinsic adsorption-activation coupling of π-backbonding theory, markedly alleviating catalyst poisoning, and further offers a cost-effective, high-efficiency paradigm for high-throughput screening of advanced catalytic materials.