Exploring the impact of operating parameters and catalyst design on propane catalytic oxidation: a machine learning perspective

Abstract

Propane is a common pollutant in the petroleum industry. In this study, machine learning coupled with SHAP (SHapley Additive exPlanations) was employed to predict the removal rate and interpret the resulting model. After collecting experimental data for low-temperature catalytic oxidation of propane using Mn-based catalysts, four machine learning algorithms and hyper-parameter optimizations were tested, yielding an accurate model that achieved a training R² of 0.999 and a test R² of 0.99. SHAP analysis showed that the ratio of surface adsorbed oxygen to lattice oxygen (O_ads/O_lat) is the primary internal factor affecting the performance of the catalyst (except for experimental conditions such as operating temperature), and this ratio has a typical volcanic relationship with the propane degradation rate. Pearson's correlation further showed a significant negative correlation between O_ads/O_lat and hydrothermal synthesis temperature (PCC = −0.38). Based on feature importance and data analysis, a series of Mn catalysts were prepared by adjusting the hydrothermal temperature, and the optimum catalyst prepared at 160 °C has a propane T₉₀ of 260 °C.