| |
Abstract |
We developed a uniform, hybrid machine learning (ML) model, integrating both supervised learning (SL) and reinforcement learning (RL), based on several datasets across different CO<sub>2</sub>and CH<sub>4</sub>conversion reactions in an atmospheric pressure glow discharge plasma, to advance plasma-based CO<sub>2</sub>and CH<sub>4</sub>conversion. Given its complex and dynamic characteristics, the SL model employs artificial neural networks (ANN) to predict performance, demonstrating excellent alignment with the entire experimental data. The RL model subsequently provides the optimization protocol, which prioritizes coarse adjustments to high-impact parameters then fine-tuning low-impact ones, to obtain the best performance. Furthermore, we also investigated the simultaneous optimization of the syngas ratio (SR) and energy cost (EC), resulting in a maximum SR of 2.12, combined with a minimum EC (syngas) of 2.04 eV per molecule (<italic>i.e.</italic>, 352 kJ mol<sup>−1</sup>), which is close to the best experimental data obtained for further methanol synthesis, when accounting for suitable weighting between SR and EC in the model. Our study emphasizes the importance of interpreting ML results based on prior knowledge and human analysis. We hope this work encourages a more critical view on the application of ML when studying plasma-based gas conversion. |
|
