Predicting hydrogenolysis reaction barriers of large hydrocarbons on metal surfaces using machine learning: Implications for polymer deconstruction

Abstract

Calculating activation energies of chemical reactions for large reaction networks is computationally demanding. Traditional Brønsted-Evans-Polanyi (BEP) relationships are useful in this regard but are prone to substantial estimation errors. Here, we explore machine learning models to predict activation energies for hydrogenolysis reactions on transition metals, including Ru, Pt, and Rh. The curated dataset includes 380 DFT-calculated activation energies of C-C and C-H scission reactions ranging from C1 to C6 species. The mean absolute error of traditional BEPs is 0.30 eV, and published BEPs on smaller hydrocarbons give even larger errors. In comparison, an XGBoost model achieves a mean absolute error of 0.19 eV for the test set without needing additional DFT-calculated energies. The model selects the homologous series among reactions and the electronegativity and atomic mass of the metal atoms as key features. The model appears transferable across metals. This approach can provide accurate and rapid estimation of activation energies of large reaction networks, such as those in plastics recycling, to accelerate catalyst screening. We showcase the impact of hydrocarbon chain length and branching on the barriers, discuss implications for the depolymerization rate of isotactic polypropylene and low vs. high-density polyethylene, and demonstrate the feasibility of hydrogenolysis catalyst screening.