Machine learning descriptors for CO activation on iron-based Fischer − Tropsch catalysts

Abstract

Due to the development of material synthesis and characterization technology, as well as limited computational resources, the understanding of CO activation on Fe-based Fischer − Tropsch synthesis (FTS) catalysts is still changing, making catalyst screening and rational design difficult. In this work, we propose a novel model that bridges the structure of common iron carbides (including o-Fe₇C₃, χ-Fe₅C₂, θ-Fe₃C, η-Fe₂C and ε-Fe_2.2C) with their CO activation capability. Using spin-polarized density functional theory (DFT), we explored CO activation pathways on a series of defective o-Fe₇C₃ surfaces. Advanced machine learning (ML) algorithms suitable for small datasets were employed to construct descriptor formulism with high predictive power for CO dissociation barriers. The ML-derived descriptor formulism unifies the catalytic expressions of various iron carbide phases, emphasizing the crucial roles of work function, carbon-vacancy formation energy, CO adsorption energy, coordination number, and the size of reaction sites in the CO dissociation process. This approach provides a deeper understanding of catalytic performance of distinct iron carbide surfaces and is applicable for designing high-performance catalysts for FTS, thereby accelerating catalyst development. Furthermore, the strategy for identifying descriptors with a limited dataset highlights the potential of combining DFT and ML methods.