Multimodal Machine Learning with 3D-Weighted-Matrix Encoding for High-Throughput Design of High-Performance Polyurethanes.

Polyurethanes (PUs) are ubiquitous in our daily life, while facing fundamental challenges in designing materials with targeted mechanical properties due to their inherent structural complexity. To address this, we developed an extensible high-throughput screening framework that combines machine learning, multimodal feature engineering, and feature fusion strategy to enable the mechanical property prediction of PU materials. Specifically, an effective 3D-Weighted-Matrix encoding method was proposed to represent polyurethane monomers, indicating better performance than conventional molecular descriptors (23% improvement in feature discriminability). Synthesis process parameters were also digitized through logic-based encoding and fused with structural features (including chemical structure representations via 3D-Weighted-Matrix and molecular descriptors as well as synthesis process information) via an early fusion architecture, yielding a multimodal deep learning model capable of concurrent prediction of Young's modulus, tensile strength, and elongation at break with mean coefficient of determination ( $R^2$ ) values exceeding 0.86. With this model, we then performed combinatorial screening of more than 150 million molecular and process combinations, identifying optimal candidates that promote various mechanical performance metrics. This work enhances our comprehension of the intrinsic structure - property correlations in PU and introduces a powerful computational framework for the accelerated development of high - performance polyurethane materials.