Electrolyte molecule design by using machine learning

Abstract

Recently high-voltage Lithium-ion batteries (LIBs) has gained a lot of attraction because of the potential application in electric vehicles However conventional carbonate solvent-based electrolytes are less stable against high-voltage cathodes In this thesis machine learning to design new electrolytes for high-voltage Lithium- ion battery have been discussed Different filters are set up for electrolyte molecular selection Among these filters the two important parameters for characterizing electrolytes are reduction potential (RP) and oxidation potential (OP) Machine learning models are trained to predict OP/RP from molecular structures Additionally we develop an algorithm which automatically modify given molecules by adding functional groups So far about 160 million distinct molecules are generated The machine learning model allows us to screen these generated molecules at a large scale and select ideal molecules Furthermore the generative model variational autoencoder was applied to carry out inverse design to generate ideal molecules The machine learning techniques lead to a new way to create brand new functional molecules for electrolyte without doing complicated calculations and costly experiments The methods could serve as a first step design towards further investigation into other necessary properties of electrolytes for high- performance LIBs

Date of Award	2020
Original language	English
Supervisor	Wen-Dung Hsu (Supervisor)