Comprehensive data mining of metal-organic framework literature using large language model

Abstract

Metal-organic frameworks (MOFs) offer several characteristics such as porosity and tunability, making them valuable in diverse applications like sensors and adsorption. For synthesizing MOFs with desired properties, it is important to find out the structure-property relationship. Up until now, we have explored this relationship through a machine learning model trained on simulation data. However, simulation data has limitations as it is derived from various assumptions. Because of this, it is necessary to train the model using experimental data, but it is difficult to collect experimental data. In this thesis, we employed a large language model for comprehensive data mining to address this issue. This approach involved simultaneously extracting overall properties and synthesis conditions from both tables and text in papers. The results confirmed a high accuracy of data mining with an F1 score of 0.9 or above.