Multi-modal cascade feature transfer for polymer property prediction

TL;DR

This work tackles data scarcity and multimodality in polymer property prediction by introducing a multi-modal cascade model that transfers features from a pre-trained GCN of molecular structures to a downstream predictor that also consumes tabular descriptors and additives. It systematically examines three transfer options based on which GCN layer is used for features and demonstrates superior data efficiency and predictive accuracy on Neat Resin and Compounding Tg datasets, often requiring far fewer samples than baseline descriptor-based models. Key contributions include validating layer-wise feature transfer within a cascade framework, showing practical data reductions (e.g., down to 14–69% of required data for comparable accuracy), and integrating compatibility with PolyCL while highlighting interpretability limitations and future directions for attention-based models. The approach offers a concrete path to leverage large-scale structural pre-trained models to accelerate polymer design under data constraints, with potential extension to multiple properties and additive encoding via LLMs.

Abstract

In this paper, we propose a novel transfer learning approach called multi-modal cascade model with feature transfer for polymer property prediction.Polymers are characterized by a composite of data in several different formats, including molecular descriptors and additive information as well as chemical structures. However, in conventional approaches, prediction models were often constructed using each type of data separately. Our model enables more accurate prediction of physical properties for polymers by combining features extracted from the chemical structure by graph convolutional neural networks (GCN) with features such as molecular descriptors and additive information. The predictive performance of the proposed method is empirically evaluated using several polymer datasets. We report that the proposed method shows high predictive performance compared to the baseline conventional approach using a single feature.

Figures (14)

• Figure 3.1: Conventional way (benchmark) and Our model
• Figure 3.2: Detailed structure of the proposed pipeline for a polymer property prediction model. First, feature extraction is performed from molecular structure data represented as graphs using a pre-trained GNN model (left side of the figure). The extracted features, along with other data such as molecular descriptors and compounding information, are then used as inputs to a machine learning algorithm to obtain predicted values for the desired polymer properties (right side of the figure).
• Figure 3.3: The distribution of glass transition temperature. The area of two target data: Neat Resin and Compound are overlapped with the are of Source data.
• Figure 4.1: Results of transferred to Neat Resin Results of transferred to target Neat resin datasets. In each plot, the horizontal axis represents the number of training data points in the target domain, while the vertical axis shows the R2 values of the resulting models. It can be observed that feature transfer from the GNN significantly improves the prediction accuracy of polymer properties across all cases. Among the three options, feature transfer from the L - 1-th layer of the GNN model (the case (b)) exhibited the highest transfer efficiency.
• Figure 4.2: Results of transferred to target compound datasets left top: (a) all features, right top: (b) selected RDKit + GCN features, left down: (c) selected RDKit + GCN features + additives, right down: (d) RDKit + features from PolyCL. In each plot, the horizontal axis represents the number of training data points in the target domain, while the vertical axis shows the $R^2$ values of the resulting models.