Machine Learning based Models for the Design of Solid Polymer Electrolytes

Abstract

With the prolific popularization and development of lithium-ion batteries, safety issues associated with the use of flammable organic electrolytes have increasingly garnered more attention. A promising alternative to organic liquid electrolytes are solid polymer electrolytes (SPEs) which demonstrate low flammability, good processability and no leakage issues.

However, presently known SPE candidates fall short of the required performance requirements, which are reliant on meeting a variety of material property requirements, such as polymer amorphicity, high ionic conductivities at room temperature, large electrochemical stability windows (4V vs Li+/Li), high Li ion transference, moderate tensile strength and thermal stability. Parsing the expansive polymer chemical space for viable SPE candidates which meet the aforementioned criteria is a non-trivial task.

My work involves the use of data-driven methods and machine learning methods to build predictive models of a variety of polymer properties relevant for the SPE application. Development of such predictive models require the collection and curation of the requisite data (from computational and experimental sources) for polymers spanning large enough chemical space, followed by the actual building of the machine learning models. These predictive models are then used to rapidly screen a large candidate space of 13,388 synthesizable polymers to identify new polymers and subsequent chemical rules for promising and reliable SPEs. Finally, we also develop models to assist in the synthesis of polymer electrolytes, specifically for identifying solvent and non-solvents for SPEs using machine learning methods to train a massive database. The work presented in this thesis highlights the ability of combining computational and experimental data with data-driven methods to accelerate the design and discovery of high-performing and reliable solid polymer electrolytes.