Predicting Polymer Flammability Using Machine Learning Methods

Abstract

Solid state batteries are a promising replacement for traditional liquid electrolyte batteries due to their superior stability and inflammability. Conventional lithium ion batteries use an organic liquid electrolyte, which is a potential hazard when it undergoes a decomposition reaction leading to an explosion. A solid, non- flammable electrolyte is a promising alternative. However, the chemical space of polymers is diverse and identifying new materials for electrolytes requires a critical screening criterion. Polymer flammability can be quantified using the limiting oxygen index (LOI), which is the minimum oxygen concentration necessary to sustain stable combustion. Polymers with an LOI greater than 21 are inflammable at room temperature and can be deemed safe for use in a battery. In this work, we predict the LOI for new polymers using a combination of data driven and machine learning methods. We have carefully curated a dataset of polymers and their associated LOI values, which are experimentally determined. A fingerprinting scheme is used to numerically represent the polymers from an atomistic to morphological length scale. We then use a gaussian progress regression model to map the polymers from their fingerprint space to their LOI. This model can then be used to rapidly predict the LOI of a polymer and its associated uncertainties. This study could rationally guide the design of solid polymer electrolytes which are thermally stable and inflammable.