COMFO: Integrated Deep Learning Model Facilitates Discovery of Multifunctional Polyimide Materials

Abstract

Rapid development in 5G and electronic appliance fields places higher demands on the dielectric, thermal and mechanical properties of materials. Dielectric materials with excellent comprehensive properties have become an urgent need of the times. Emerging machine learning techniques could greatly accelerate the discovery of high-performance dielectric materials. However, it remains unknown whether traditional 2D fingerprints or descriptors can extract molecular structure information more completely. In this study, data on four types of properties of polyimide (PI), including dielectric constant, glass transition temperature, tensile modulus and coefficient of thermal expansion, were collected to construct a deep learning model-COMFO to explore polyimide dielectric materials with excellent comprehensive performance. Our COMFO model could extract the key feature information in the molecule from three perspective learning tasks as well as process and learn them. Specifically, the three learning tasks include extracting the feature information in the SMILES sequence using a large language model, the bidirectional encoder Transformer; extracting the information about the atoms and bonds of polymer molecules from molecular graph using the Attentive FP network; and extracting the information about the substructures of polymer molecules through molecular fingerprints. The multi-perspective feature extraction task gave our model a more excellent performance (R²>0.90). The performance of the model was confirmed by various ways, including experimental validation, MD simulation validation, and comparison with 12 other models. Design guidelines for low dielectric constant PIs were discovered by monomer structure analysis. High-throughput virtual screening of 158,022 unknown PIs was performed and three PIs with excellent comprehensive properties (especially dielectric properties) were identified. MD and DFT approaches verified and analyzed the properties of these three potential high-performance PIs. In the future, this research could also contribute to the forward development of materials in other fields.