Multimodal Modeling for Polymer Property Prediction and Decoupling of Structure-Property Relationship.

The key challenge in the development of a target property polymer stems from the unclear structure-property relationship (SPR). A multimodal graph neural network (GNN) framework, GeoALBEF, is proposed to address this problem. GeoALBEF introduces a training process based on an "Align before Fuse" architecture that optimizes a three-stage loss, which deeply fuses polymer graph information and text information. Benchmark tests constructed based on 5,000 polymer samples showed that GeoALBEF reduced the mean relative RMSE by 8.6% compared to the suboptimal model in the prediction task of 24 key properties in 6 categories. It is especially worth pointing out that the model achieves functional group level interpretability through the attention mechanism and the functional group averaging strategy. The interpretability was mapped using visualization methods. Additionally, we used reinforcement learning to quantify this interpretability and successfully decoupled the SPR of the polymers. This multimodal system, which combines high-precision prediction capability and mechanism analysis, establishes an intelligent mapping paradigm from polymer structure to property and is expected to promote the accelerated optimization of target property polymers.