| Abstract Scope |
As renewable energy sources are increasingly gaining traction as an alternative to scarce fossil fuels, exploratory research on energy storage devices, particularly Li-ion batteries (LIBs), has become prolific. To safely utilize a high-performing LIB, a few alternatives have been suggested, one of which is replacing the liquid electrolyte with a suitable solid polymer electrolyte (SPE). SPEs offer several advantages over conventional liquid electrolytes, viz., low flammability, good processability, and no leakage issues. SPEs also eliminate the need for a separator, thereby decreasing the chances of an accident. A promising SPE candidate should have high Li-ion conductivity, a low glass transition temperature - both of which depend on the degree of crystallinity of the polymer.
Data-driven efforts to predict the polymer crystallinity have been scarce. In this first-of-its-kind work, we develop a multi-fidelity dataset of over 400 polymers which comprises of a high-fidelity dataset which uses explicit experimental crystallinity information and another low-fidelity dataset which uses theoretical group contribution methods with experimental data. With this dataset, we then develop and compare machine learning models viz. conventional gaussian process regression with the high-fidelity dataset and co-kriging with the multi-fidelity dataset. Through this effort we aim to predict the polymer crystallinity of new polymers instantly and also use it as a screening criterion to design new materials for solid polymer electrolytes. |