Machine Learning-Assisted Multi-Objective Optimization for Rational Design of Perovskite Catalysts for VOCs Removal

Exponential emission of volatile organic compounds (VOCs) from fuel processing and fossil fuel combustion has led to significant environmental challenges. Catalytic conversion using perovskite oxides has emerged as a promising approach for VOCs removal. However, the catalytic performance of perovskites is influenced by complex and nonlinear interactions between composition, structure, and synthesis conditions, which makes predicting catalytic activity and designing high-performance catalysts inefficient. To address this challenge, this paper proposes a machine learning framework that incorporates three components: (i) data-driven modeling for catalytic activity prediction, (ii) feature analysis for mechanistic interpretation and interaction elucidation, and (iii) multiobjective optimization for composition design under stoichiometric constraints. An experimental database was constructed from existing research, focusing on activation energy ( E a ), T 50%, and T 90% as targets. The extremely randomized trees (extra-trees) model was selected for optimal prediction performance. Feature importance analysis identified Co as the most critical factor, followed by La, Fe, Sr, pore volume, and calcination temperature. Further analysis revealed that Co content, calcination time, and specific surface area exhibit strong positive effects, while Mn content, toluene concentration, and pore diameter exhibit negative effects beyond optimal ranges. Notably, Co content exhibits synergistic interactions with pore volume and calcination time, significantly influencing catalytic performance. Under A-site vacancy and B-site stoichiometric constraints, the NSGA-II generated a Pareto front that reveals the internal competitive relationship between E a, T 50%, and T 90% . Using CRITIC weighting and TOPSIS, the optimal composition La 0.84 Sr 0.08 Co 0.82 Fe 0.18 O 3 was identified. Compared with database averages, this composition reduces E a, T 50%, and T 90% by 37.3%, 3.0%, and 12.1%, respectively. This work introduces a transferable framework combining data modeling, mechanistic explanation, and multiobjective optimization for the rational design of high-performance perovskite catalysts.