| 초록 |
An efficacy of liquid organic hydrogen carrier (LOHC) system is mainly determined by the properties of the materials. The dominant factors are the following properties; 1) hydrogen storage capacity, 2) dehydrogenation energy (DE), 3) liquid state range. We have performed a screening study using these three factors through machine learning (ML) approach. We constructed the training datasets of DE calculated from DFT calculations (Lab own dataset) and of melting point & boiling point from EPI PHYSPROP dataset. Then, a machine learning structure was optimized based on Graph SAGE model. The mean absolute errors of each prediction model were estimated as 3.73 kJ/molH2, 14.73 ℃, and 38.62 ℃, respectively. This error is relatively lower than other graph-based ML models. We expanded an additional predictive dataset for the material screening, and predicted the LOHC properties by the ML models. As a result, we confirmed that the prediction values fit the DFT results quite well, which is useful to screen the LOHC materials. Our study gives an insight for material design through in silico ML method. |
