High-throughput screening of stable Ag–Pd–F catalysts for formate oxidation reaction using machine learning

Abstract

The Ag–Pd–F materials demonstrates excellent catalytic performance as formate oxidation reaction (FOR) catalysts in direct formate fuel cells (DFFCs). The development of efficient FOR catalysts has been hindered by the lack of a systematic approach to catalyst design. Traditional methods, which rely on experimental trial-and-error and computationally intensive density functional theory (DFT) calculations, are not only costly but also time-consuming. The two crystal graph convolutional neural networks (CGCNN-1 and CGCNN-2 models) were developed utilizing the formation energy and convex hull distance as targets, respectively, to predict the novel Ag–Pd–F catalytic materials in this study. From an initial set of 20 130 alternative Ag–Pd–F structures, the CGCNN-1 and CGCNN-2 models identified 728 and 259 potentially stable Ag–Pd–F materials with the predicted convex hull distance less than zero. 149 metastable Ag–Pd–F candidate materials with the calculated convex hull distances less than 100 meV per atom and 8 novel low-energy stable Ag–Pd–F compounds with the calculated convex hull distances less than zero were validated using DFT calculations. The validated novel low-energy stable compounds include Ag₂PdF₆_La₂WO₆, Ag₂PdF₆_Na₂PdF₆, AgPd₂F₁₂_CaCr₂F₁₂, Ag₂PdF₆_Sm₂WO₆, AgPd₂F₆_Ca₂H₆Os, Ag₂PdF₆_Ni(IO₃)₂, Ag₃PdF₂₀_BiSb₃F₂₀ and AgPd₃F₂₀_BiSb₃F₂₀. Among these compounds, Ag₂PdF₆_La₂WO₆, Ag₂PdF₆_Na₂PdF₆ and Ag₂PdF₆_Ni(IO₃)₂, with convex hull distances of −20.42, −19.78, and −17.44 meV per atom, respectively, also exhibit dynamic stability. Ag₂PdF₆_La₂WO₆ (100) and Ag₂PdF₆_Na₂PdF₆ (100) facets demonstrated all downhill pathways in free energy diagram for the FOR, indicating that these two facets can thermodynamically catalyze the spontaneous oxidation of HCOO⁻ to CO₂ and H⁺. These findings provide valuable insights for designing new catalysts for FOR in DFFCs and demonstrate the effectiveness of input selection in machine learning models for materials discovery.