Enhancing Catalyst Performance Prediction with Hybrid Quantum Neural Networks: A Comparative Study on Data Consistency Variation

Abstract

Data consistency affects the robustness of machine learning-based models. Most experimental and industrial data have low consistency, leading to poor generalization performance. In this study, a hybrid Quantum Neural Network (hybrid QNN) with superior generalization capabilities, was compared with established machine learning models, including artificial neural networks and decision-tree-based methods such as CatBoost and XGBoost. We evaluated these models by predicting the catalyst performance across different data-consistency scenarios using two catalyst data sets: a low-consistency preferential oxidation of CO (PROX) catalyst and a high-consistency oxidation coupling of methane (OCM) catalyst. The hybrid QNN performed better in both low- and high-consistency environments, demonstrating robust generalization capabilities. In the regression tasks, the hybrid QNN achieved a 6.7% lower mean absolute error (MAE) for the PROX catalyst and a 35.1% lower MAE for the OCM catalyst compared with the least-performing model. Adaptability is crucial in catalysis, where data scarcity and variability are common. Our research confirms the potential of the hybrid QNN as a comprehensive tool for advancing catalyst design and selection by achieving high accuracy and predictive power under diverse conditions.