Polymer thermal properties, including glass transition (Tg), melting (Tm), decomposition (Td), and crystallization (Tc) temperatures, alongside the softening point (SP), are pivotal for material design yet challenging to determine efficiently via laborious experiments or computationally intensive simulations. To leverage the complementary strengths of diverse molecular representations, this work introduces a deep learning framework featuring a gated fusion mechanism that integrates molecular graph representations and hybrid fingerprint descriptors. Evaluation on five thermal properties demonstrates that the proposed model shows improved performance over traditional benchmarks in single-task learning. Ablation studies and gating mechanism analysis reveal that the model adaptively prioritizes graph or fingerprint features, enabling effective multimodal fusion. Furthermore, a multitask learning strategy exploits latent correlations to reduce prediction errors for properties with limited data, offering an efficient and unified framework. This dual approach provides a competitive tool for accelerating data-driven material discovery.
Lin et al. (Mon,) studied this question.