Multi-Property Prediction of Polymers and Exploration of Optimal Polymer Structures with Deep Learning

Polymer is used in every industrial field these days due to its workable, lightweight characteristics and flexibility that it can be composed with other materials to give additive properties. However, it is difficult to find high-performance polymers to satisfy all demanded property for specific products by experiments only. Recently, Materials informatics (MI) which is a data-driven material development has grown thanks to the innovation of computing technology and the improvement of databases, and it enables us to predict some properties of materials without experiments. Moreover, machine learning (ML), especially deep learning (DL) which uses deep neural networks (DNN) that mimic the function of brain’s biological neurons, helps us to find high-level correlations and features between inputs and outputs from massive and complex datasets in a relatively short time.<br/>At first, we developed DNN to predict multi-property of polymers listed in PoLyInfo. In previous studies, only Morgan fingerprint was considered as inputs of prediction networks because the fingerprint consists of only 0 and 1 and they can be input to networks directly. But considering solving an invert problem after predicting properties, it is inconvenient that the input of prediction networks is the fingerprint because the molecular structure cannot be reconstructed completely with its fingerprint. For this reason, we designated the input of DNN as the string of the canonical SMILES (simplified molecular input line entry system) of a polymer which describes the structure of a molecule in our study. The DNN has LSTM (long short-term memory), which can deal with characters of SMILES through vectorization, and a dense layer whose input is Morgan fingerprint converted from SMILES in order to use both microscopic and macroscopic information of the structure. We predicted electrical, thermal and mechanical properties with DL and obtained enough accuracies to apply the DNN to the inverse problem to estimate optimal molecular structures that satisfy multiple demanded properties.<br/>After evaluating the performance of the DNN, we explored optimal molecular structures by solving the invert problem with the DNN we have developed. Firstly, we developed another network that was trained to generate correct SMILES using GAN (generative adversarial networks) and some novel methods. After that, we adopted BO (Bayesian optimization), GRP (Gaussian Process Regression) and some optimization methods to obtain the high-performance polymers and compared the effectiveness of these methods by evaluating the properties of suggested polymers. As an example of insulated cables which are required all electrical, thermal and mechanical properties, we could obtain optimal polymer materials for the cables by performing a series of DL calculations.