A method for optimizing catalyst preparation conditions based on machine learning and genetic algorithm

Abstract

ABSTRACT Multiphase catalytic oxidation is effective for volatile organic compound (VOC) treatment, but traditional catalyst optimization suffers from inefficiency and local optima. This work proposes an integrated strategy for Mn-Ca/γ-Al 2 O 3 catalysts that targets xylene breakdown by combining machine learning (ML) and genetic algorithms (GA). Orthogonal and single-factor experiments were used to create a high-quality dataset of 225 samples (45 real measured experimental points and 180 interpolation-augmented synthetic samples) under fixed reactor-operating conditions (catalytic packing mass = 90 g, ROS injection rate = 60 mg/L min), covering four adjustable catalyst-preparation parameters: active doping loading, Ca/Mn ratio, calcination temperature, and duration. Gradient Boosting Regression (GBR) and Random Forest (RF) models were developed; the RF model had the best prediction accuracy (test set R-squared (R²) = 0.949, Mean Absolute Error (MAE) =0.022). The SHapley Additive exPlanations (SHAP) value analysis reveals that the Ca/Mn molar ratio is the dominant factor contributing 43.94%, followed by calcination temperature (21.96%). GA adjusted settings (5.4% active doping loading, Ca:Mn = 1:4, 423 °C, 3.75 h) using the RF model as the fitness function produced a xylene removal rate of 93.7%. This ML-GA framework provides a green, low-carbon paradigm for catalyst development in heterogeneous catalysis.