Large-Scale Glass-Transition Temperature Prediction with an Equivariant Neural Network for Screening Polymers

The practically infinite chemical and morphological space of polymers makes them pervasive with applications in materials science but challenges the rational discovery of new materials with favorable properties. Polymer informatics aims to accelerate materials design through property prediction and large-scale virtual screening. In this study, a new method (Lieconv-Tg) has been developed to predict glass-transition temperature (Tg) values from repeating units of polymers based on Lieconv, which is equivariant with transformations from any specified Lie group. The introduction of equivariance allows the prediction of molecular properties from their 3D structures, independent of orientation and position. A total of 27,659 homopolymers with Tg values were collected from PolyInfo, and a standard data set containing 7166 polymers (named data set_Tg) was created for training a robust Lieconv-Tg model. Using the 3D coordinates as input, Lieconv-Tg performs better than Edge-Conditioned Convolution (ECC), and the mean absolute error (MAE) is significantly reduced by similar to 6 from similar to 30 to similar to 24 on both the validation set and the test set, and the R-2 value for both the validation set and the test set can reach 0.90. Lieconv-Tg is thus used to screen promising candidates from a benchmark database named PI1M with 995,800 generated polymers. However, there are some implausible repeating units in PI1M. To get more reasonable candidates from PI1M, a new filtering method has been accomplished by utilizing Morgan fingerprints at the polymerization points (MF@PP) of repeating units in data set_Tg. The combination of a standard data set, Lieconv-Tg, and a more reasonable screening strategy provides new directions in materials design for polymers.