Machine learning discovery of high-temperature polymers

Abstract

To formulate a machine learning (ML) model to establish the polymer's structure-property correlation for glass transition temperature $T_g$ , we collect a diverse set of nearly 13,000 real homopolymers from the largest polymer database, PoLyInfo. We train the deep neural network (DNN) model with 6,923 experimental $T_g$ values using Morgan fingerprint representations of chemical structures for these polymers. Interestingly, the trained DNN model can reasonably predict the unknown $T_g$ values of polymers with distinct molecular structures, in comparison with molecular dynamics simulations and experimental results. With the validated transferability and generalization ability, the ML model is utilized for high-throughput screening of nearly one million hypothetical polymers. We identify more than 65,000 promising candidates with $T_g$ > 200°C, which is 30 times more than existing known high-temperature polymers (∼2,000 from PoLyInfo). The discovery of this large number of promising candidates will be of significant interest in the development and design of high-temperature polymers.

Keywords: feature representation; glass transition temperature; high-throughput screening; machine learning; polymer.

Graphical abstract

Figure 1

Chemical space visualization of dataset-1, dataset-2, and dataset-3 (A) 2D visualization based on descriptors and fingerprints using the t-SNE algorithm. Dataset-1 has reported $T_g$ values, and each data point is colored based on the corresponding $T_g$ value. Dataset-2 and dataset-3 do not have reported $T_g$ values, colored with yellow and red, respectively. (B) Set diagram showing representative substructures in dataset-1 (green circle), dataset-2 (yellow circle), and dataset-3 (red circle) based on Morgan fingerprint. Some substructures are common for all datasets, while some others are unique to certain datasets.

Figure 2

Three types of feature representation calculated based on the polymer's SMILES notation for ML models: molecular descriptor, Morgan fingerprint, and image

Figure 3

Performance of four ML models (A) The Lasso regression model using descriptors as input features (Lasso_Descriptor model). (B) The Lasso regression model using fingerprints as input features (Lasso_Fingerprint model). (C) The DNN model using fingerprints as input features (DNN_Fingerprint model). (D) The CNN model using images as input features (CNN_Image model). (E) The comparison between the MD-simulated $T_g$ and the ML-predicted $T_g$ on 20 polymers randomly selected from dataset-2. Three dashed lines are a unity line and lines with a mean absolute error of 40°C. The chemical structure of these 20 polymers is followed by their MD-simulated $T_g$ value.

Figure 4

Substructures with the highest absolute weight based on Morgan fingerprint and Lasso ML model The central atom of the substructures is highlighted in blue. Aromatic atoms are highlighted in yellow. Connectivity of Atoms is highlighted in light gray.

Figure 5

High-throughput screening of high $T_g$ polymers with the DNN_Fingerprint model The $T_g$ distribution of the dataset-1, dataset-2, and dataset-3 are plotted in green, yellow, and red, respectively. The polymer samples on the right are following by their predicated $T_g$ and true $T_g$ values. For the sample in dataset-1 (green box), true $T_g$ is the collected experimental value. For the samples in dataset-2 (yellow box) and dataset-3 (red box), true $T_g$ is the MD-simulated value. More than 1,000 real polymers and 65,000 hypothetical polymers were discovered with $T_g$ > 200°C.

Figure 6

Comparison of key substructures and functional groups in high-$T_g$(>200°C) polymers (A) Comparison of the 18 substructures recognized in Figure 4. (B) Comparison of the six high-$T_g$-related functional groups recognized in Table 5.