Autoencoder Derived Datasets for Multi-Layer Perceptron Development of Methanol Reforming Catalysts

Abstract

In recent years, an increasing number of studies have employed machine learning for catalyst development to analyze large datasets and support decision-making. However, real experimental datasets often exhibit various data quality issues, with models frequently suffering from low accuracy due to outliers. To tackle this challenge, this study established a machine learning framework combining an autoencoder (AE) with a multi-layer perceptron (MLP) using methanol reforming as a case study. The AE component ensured data quality by intelligently identifying and removing outliers, thereby improving the performance of the subsequent models. The MLP captured the complex correlations, enabling accurate predictive modeling. SHapley Additive exPlanations (SHAP) analysis provides optimal parameter ranges for features across different catalyst systems, rather than merely revealing overall trends. This study provides valuable insights into improving the input data quality for machine learning models and paves the way for connecting artificial intelligence and catalyst design.