Machine learning assisted rediscovery of methane storage and separation in porous carbon from material literature

Porous carbon (PC) has been widely regarded as one of the most promising absorbents for methane storage. Studies show that its uptake capacity and selectivity highly depend on textural structures. Although much effort has been made, unveiling their detailed structure-performance relationship remains a challenge. Here, we propose an innovative study where, with the assistance of machine learning, the hidden relationship of the textural structures of PC with the methane uptake and separation can be derived from existing data in material literature. Machine learning models were trained by the data, including specific surface area, micropore volume, mesopore volume, temperature, and pressure as the input variables and methane uptake as the output variable for prediction. Among the tested models, the multilayer perceptron (MLP) shows the highest accuracy in predicting the methane uptake. In addition, the model enables to automatically construct a uptake performance map in terms of micropore volume and mesopore volume. The obtained MLP model was also extended to explore the CO2/CH4 selectivity by retraining it with the data collected from literature of PC for the CO2 uptake. The constructed 2D selectivity map shows that the high selectivity can be achieved in the low CH4 uptake region.

Automated summary

Porous carbon is considered a promising absorbent for methane storage, with its uptake capacity and selectivity dependent on textural structures. Utilizing machine learning, a model was able to predict methane uptake based on specific surface area, pore volume, temperature, and pressure, and construct a map of uptake performance. The model was also extended to explore CO2/CH4 selectivity in the low CH4 uptake region.