Enhanced T g Prediction in Polyimide via PolySDA: A Novel Shallow‐Deep Multimodal Fusion Framework

Polyimide, as a specialized engineering material, is widely used in aerospace, electronic packaging, and high-temperature coatings. Traditionally, determining the physical properties of polyimide (such as the glass transition temperature) involves expensive experimental equipment, leading to a cumbersome and costly process. Although machine learning techniques have recently been employed for property prediction, most approaches rely on single-modal representations, overlooking the fact that molecules often exhibit multiple modes of representation. While a few studies have explored multimodal fusion, they have not fully accounted for the potential impact of shallow-level features on predictive performance. In response to these challenges, a novel multimodal algorithmic framework-PolySDA (Polyimide Shallow-Deep Alignment Framework)-is proposed. This framework jointly exploits and aligns both shallow and deep multimodal features of molecules, thereby enhancing prediction accuracy. PolySDA introduces specialized modules in its front-end and back-end stages to maintain consistent feature shapes and facilitate similarity calculations, coupled with a dedicated loss function to achieve progressive alignment of shallow and deep representations. Experimental results on a polyimide dataset indicate a notable improvement in predictive performance, confirming the effectiveness of the proposed approach.