Machine learning-based prediction and optimization of plasma-catalytic dry reforming of methane in a dielectric barrier discharge reactor

We developed an innovative machine learning (ML) model, including a supervised learning (SL) and reinforcement learning (RL) model, to predict and optimize the plasma-catalytic dry reformation of methane (DRM) over Ni/Al2O3 catalysts in a dielectric barrier discharge (DBD) reactor based upon experimental data. To tackle its intricate and non-linear characteristics, the SL model uses artificial neural networks (ANN) to accurately predict the performance, achieving excellent consistency with the experimental results. The RL model subsequently investigates the optimal optimization policy, namely starting with a coarse tuning of the more influential parameters, followed by fine-tuning of the less important parameters, to obtain the best performance. The optimal results show that a discharge power at lowest bond (i.e., 20 W) but CO2/CH4 ratio at highest bond (i.e., 1.5) result in the minimum energy cost (21 eV/molec), validated by our SL model and experimental data. Furthermore, we also investigated the simultaneous optimization of total conversion and energy cost, resulting in a maximum total conversion of 36 %, combined with a minimum energy cost of 34 eV/molec, at a Ni loading of 9.5 wt%, discharge power of 60 W, and total flow rate of 74 mL/min. Our ML model showcases an impressive capacity to derive advantageous insights from existing datasets, thereby advancing and optimizing plasma- catalytic chemical processes.