Jin Hoon Yang1,Ji Young Lee1,Eun Ho Sohn1,Hyunju Chang1,Seunghun Jang1
Korea Research Institute of Chemical Technology1
Jin Hoon Yang1,Ji Young Lee1,Eun Ho Sohn1,Hyunju Chang1,Seunghun Jang1
Korea Research Institute of Chemical Technology1
In this study, we propose a novel machine learning approach to accurately predict fluoropolymers' glass transition temperature (T<sub>g</sub>) and demonstrate its potential in guiding the design of high T<sub>g</sub> copolymers. Firstly, we utilize the QM9 dataset for model pre-training, providing robust molecular representations for subsequent transfer learning on a specialized copolymer dataset. Our pre-trained model expertly encodes complicated molecular structures and general molecular properties using atom-level (graph) and global molecular-level (global state) features. This extensive feature set is processed via a dual network system, with the outputs merged to form a comprehensive molecular descriptor. Dealing with a small copolymer dataset, we encounter significant discrepancies between individual models, a common issue in limited data. We address this problem by adopting an ensemble approach and providing more reliable and robust predictions. Finally, we can navigate a vast chemical space comprising 61 monomers and identify promising candidates for developing high T<sub>g</sub> fluoropolymers. Our work shows the potential of machine learning in materials design and discovery and the effectiveness of ensemble models in addressing the challenges associated with small datasets.