Metaheuristics-guided active learning for optimizing reaction conditions of high-performance methane conversion

Abstract

Converting greenhouse gases into value-added chemical compounds has been widely studied in chemical science for sustainable industry. In particular, nonoxidative coupling of methane (NOCM) that transforms methane into useful chemical compounds has received great interests because methane is a naturally abundant greenhouse gas. NOCM has received significant attention in sustainable industry for addressing the climate crisis. However, there are complex optimization problems in maximizing the efficiency of the nonoxidative methane conversion. Although conventional data-driven methods have been successfully applied to various engineering problems, a data-driven optimization of NOCM remains a challenging problem because expensive chemical experiments should be performed to collect prior data for model training. To avoid the expensive costs of chemical experiments, we propose an active learning method that performs training data augmentation and model re-training without pre-defined unlabeled experimental data. To this end, we combine active learning with metaheuristic algorithms to perform active learning with statistically augmented data. We applied the proposed method to a high-throughput screening task to discover new reaction conditions of high-performance NOCM, and the high-throughput screening error was significantly reduced by 69.11%.