Accelerating amorphous polymer electrolyte screening by learning to reduce errors in molecular dynamics simulated properties

Abstract

Polymer electrolytes are promising candidates for the next generation lithium-ion battery technology. Large scale screening of polymer electrolytes is hindered by the significant cost of molecular dynamics (MD) simulation in amorphous systems: the amorphous structure of polymers requires multiple, repeated sampling to reduce noise and the slow relaxation requires long simulation time for convergence. Here, we accelerate the screening with a multi-task graph neural network that learns from a large amount of noisy, unconverged, short MD data and a small number of converged, long MD data. We achieve accurate predictions of 4 different converged properties and screen a space of 6247 polymers that is orders of magnitude larger than previous computational studies. Further, we extract several design principles for polymer electrolytes and provide an open dataset for the community. Our approach could be applicable to a broad class of material discovery problems that involve the simulation of complex, amorphous materials.

Early prediction of ion transport properties in solid polymer electrolytes using machine learning and system behavior-based descriptors of molecular dynamics simulations

“…ML models have also been used to reduce computational cost of MD simulations in various areas of materials science, including glass design [45,46], force field development [47,48], and catalyst discovery [49,50]. Specifically, in SPE simulations, methods such as machine learning assisted coarse-graining of MD simulations [17], learning random and systematic errors from short and long length simulations [20], and unsupervised classification of ion solvation sites [51] have been used to leverage the enormous data generated from MD simulations and increase the efficiency of these computations. Further, ML techniques have been employed to make more accurate estimations of various SPE properties such as mechanical properties (elastic modulus and yield strength) [52], ionic transport properties (ionic diffusivity, conductivity, and transference number) [17,20,53,54], and solvation environment of ions [51,53,55], from atomistic simulations.…”

“…This means the time evolution of properties serve as important system behavior-based descriptors for predicting ion transport properties. These descriptors have been employed in previous studies of SPEs [20], and their usage is still a growing area of study. However, due to the insufficient MD simulation data for SPEs, ML prediction of SPEs properties from MD simulations had been limited to a small number of polymer structures up until recently [20,54].…”

“…These descriptors have been employed in previous studies of SPEs [20], and their usage is still a growing area of study. However, due to the insufficient MD simulation data for SPEs, ML prediction of SPEs properties from MD simulations had been limited to a small number of polymer structures up until recently [20,54]. On the other hand, the more expensive behavior-based descriptors extracted from MD simulations include information about interactions between cations and anions and the time evolution of transport properties; thus, their prediction performance can be improved at the cost of longer simulations.…”

“…Currently established MD simulation workflows for polymer systems require multiple heating and cooling cycles to fully relax the structure, which make them computationally expensive [20,[25][26][27][28]. Even after a system has been fully equilibrated, the slow inter-chain hopping of Li ions [19,23] requires long simulation times to correctly extract diffusivity and conductivity.…”

“…Even after a system has been fully equilibrated, the slow inter-chain hopping of Li ions [19,23] requires long simulation times to correctly extract diffusivity and conductivity. As a result, long production runs in MD simulation (on the orders of 10s to 100s of ns) are required to ensure accurate and converged computation of ionic transport properties [19,20,29,30]. MD simulations, therefore, must be significantly accelerated to address the demand for the computational discovery of new SPEs.…”