Modeling and Predicting the Glass Transition Temperature of Vinyl Polymers by Using Hybrid Pso-SVR Method

Abstract

Based on four physicochemical descriptors (the rigidness descriptor R OM resulted by hydrogenbonding moieties group and/or rings, the chain mobility n, the molecular average polarizability and the net charge of the most negative atom q À Þ derived from the polymers' monomers structure, the support vector regression (SVR) approach combined with particle swarm optimization (PSO), is proposed to construct a model for prediction of the glass transition temperature T g of three classes of vinyl polymers, including polystyrenes, polyacrylates and polymethacrylates. The mean absolute error (MAE ¼ 13:68 K), mean absolute percentage error (MAPE ¼ 4:22%) and correlation coe±cient (R 2 ¼ 0:9252) calculated by SVR are superior to those (MAE ¼ 16:74 K; MAPE ¼ 5:30% and R 2 ¼ 0:9059) achieved by S-SAR model, and (MAE ¼ 16:83 K; MAPE ¼ 5:27% and R 2 ¼ 0:9057) achieved by ANN model for the identical training set (124 vinyl polymers), whereas the MAE ¼ 15:09 K; MAPE ¼ 4:82% and R 2 ¼ 0:9253 calculated by SVR are also better than those of MAE ¼ 17:96 K; MAPE ¼ 5:94% and R 2 ¼ 0:8952 achieved by S-SAR, and MAE ¼ 16:603 K, MAPE ¼ 5:4% and R 2 ¼ 0:9120 achieved by ANN for the same 68 test samples. Furthermore, the MAE, MAPE and R 2 for an independent set (10 vinyl polymers) predicted by SVR also reached 14.132 K, 4.25% and 0.9475, respectively. The results strongly support that the comprehensive modeling and prediction ability of SVR model surpass those of S-SAR and ANN models by applying identical training, test and independent samples. It is demonstrated that the established SVR model is more suitable to be used for prediction of the T g values for unknown vinyl polymers possessing similar structure than S-SAR model or ANN model.

Data‐Driven Materials Innovation and Applications

“…Their SVM simulation demonstrated that the concrete ultrasonic pulse velocity was the most effective parameter in improving the accuracy of estimations. SVM models have been used to predict various other material properties such as ionic conductivities, [ 106,107 ] glass transition temperatures, [ 108–110 ] catalyst active sites [ 111 ] and adsorption energies, [ 112 ] and various other properties of innovative materials. [ 113,114 ] $$\[\left[ {\frac{1}{n}\mathop \sum \limits_{i = 1}^n \max \left( {0,1 - {y_i}\left( {{\beta ^T}{x_i} - b} \right)} \right)} \right] + \lambda \parallel \beta {\parallel ^2}\]$$…”

“…Key to the efficient use of ML in the field of chemical materials is the “descriptor selection” tool, which takes the entire descriptor set as an input, or combines it into a new reduced, but more reliable, descriptor set through correlation analysis while providing a mapping to a key performance indicator (KPI) fingerprint. [ 50 ] In this section, the strategy of transforming material data to ML through descriptors is introduced; descriptors can be divided into five main types: constitutional descriptors; [ …”

“…Their SVM simulation demonstrated that the concrete ultrasonic pulse velocity was the most effective parameter in improving the accuracy of estimations. SVM models have been used to predict various other material properties such as ionic conductivities, [106,107] glass transition temperatures, [108][109][110] catalyst active sites [111] and adsorption energies, [112] and various other properties of innovative materials. [113,114] 1 max 0,1…”

Surrogate Model via Artificial Intelligence Method for Accelerating Screening Materials and Performance Prediction

“…[ 12 ] Support vector machine (SVM), which is a widely used machine learning algorithm, was used to predict ionic conductivity [ 13 ] and glass transition temperatures. [ 14–16 ]…”

Predicting the onset temperature (Tg) of Ge Se1− glass transition: a feature selection based two-stage support vector regression method

“…They proved that both the predictive accuracy and generalization ability of SVR are superior to those of the quantitative structure-property relationship (QSPR) model [39]. In 2013, Pei et al [40] used the particle swarm optimization-support vector regression (PSO-SVR) method to predict the T g of three classes of vinyl polymers, and found that PSO-SVR achieves better perfor-mance than the spectral structure-activity relationship analysis and ANN. Furthermore, Alzghoul et al [41] used several machine learning methods including MLR, partial least-squares, principal component regression, ANN and SVR, for the T g prediction of drugs, among which SVR gives the best result with RMSE of 18.7 K. Above all, machine learning based methods have achieved good prediction performance for T g values of many kinds of organic glasses, but few references on using machine learning methods to predict the T g of inorganic glasses are located.…”