Predicting the Glass Transition Temperature of Biopolymers via High-Throughput Molecular Dynamics Simulations and Machine Learning

Abstract

Nature has only provided us with a limited number of bio-based and biodegradable building blocks. Therefore, the fine tuning of the sustainable polymer properties is expected to be achieved through the control of the composition of bio-based copolymers for targeted applications such as cosmetics. Until now, the main approaches to alleviate the experimental efforts and accelerate the discovery of new polymers have relied on machine learning models trained on experimental data, which implies an enormous and difficult work in the compilation of data from heterogeneous sources. On the other hand, molecular dynamics simulations of polymers have shown that they can accurately capture the experimental trends for a series of properties. However, the combination of different ratios of monomers in copolymers can rapidly lead to a combinatorial explosion, preventing the investigation of all possibilities via molecular dynamics simulations. In this work, we show that the combination of machine learning approaches and high-throughput molecular dynamics simulations permits to quickly and efficiently sample and characterize the relevant chemical design space for specific applications. Reliable simulation protocols have been implemented to evaluate the glass transition temperature of a series of 58 homopolymers, which exhibit a good agreement with experiments, and 488 copolymers. Overall, 2,184 simulations (4 replicas per polymer) were performed, for a total simulation time of 143.052 µs. These results, constituting a dataset of 546 polymers, have been used to train a machine learning model for the prediction of the MD-calculated glass transition temperature with a mean absolute error of 19.34 K and a R2 score of 0.83. Overall, within its applicability domain, this machine learning model provides an impressive acceleration over molecular dynamics simulations: the glass transition temperature of thousands of polymers can be obtained within seconds, whereas it would have taken node-years to simulate them. This type of approach can be tuned to address different design spaces or different polymer properties and thus have the potential to accelerate the discovery of new polymers.