Elucidating the Role of Nickel Catalyst Properties, Synthesis Methods, Process Variables, and Reactor Geometry on Hydrogen Production from Ammonia

Abstract

This study explores seven machine learning (ML) models to predict the performance of cost-effective Ni-based catalysts and assess the effects of various parameters on ammonia decomposition. A comprehensive database consisting of 6447 datapoints, with 16 input features describing catalyst composition, synthesis, reaction conditions, reactor geometry, and output ammonia conversion, was collated from the literature. Among the seven ML models investigated in the study, the Artificial Neural Network(Multilayer Perceptron) (ANN(MLP)) achieved superior performance, with R2 and standardized RMSE values of 0.978 and 0.149, respectively. This was followed closely by Light Gradient Boosting (LGBM) and Extreme Gradient Boosting (XGB), with R2 values of 0.972 and 0.970, respectively. While ANN(MLP), XGB, and LGBM showed consistently high performance across 5-fold and the leave-one-paper-out cross-validation, ANN(MLP) showed exceptional performance on structurally homogeneous data, and LGBM showed enhanced generalization on completely unseen data. SHapley Additive exPlanations (SHAP) analysis identified reaction temperature and gas hourly space velocity as major descriptors to ammonia conversion over Ni-based catalysts. The two factors collectively account for 57.18% of the output. The developed ML framework can effectively map highly dimensional data and provide a reliable methodology for sustainable catalyst design.