Applying transfer learning with convolutional neural networks to identify novel electrolytes for metal air batteries

Abstract

The formation of dendrites on the anode surface of metal air batteries, such as the Li-air battery, causes significant decreases in performance over the lifetime of the battery and poses safety concerns due to short circuiting. Predictive computational methods are used to investigate novel electrolyte materials which will reduce dendrite growth, in particular liquid crystal materials for Li-air electrolytes. The literature on liquid crystal electrolytes was surveyed for materials data and used to develop a training set of liquid crystal compounds. These compounds were then used as a knowledge base to develop property structure relations for the clearing and melting temperature for liquid crystals, which are a critical design parameter for the design of novel electrolytes. This was accomplished using standard artificial neural networks trained on molecular fingerprints and convolutional neural networks (CNNs) trained on compound images. Transfer learning was also demonstrated to boost predictive performance through pre-training neural networks on larger pre-existing compound datasets. The results show that CNNs achieve comparable accuracy compared to molecular fingerprints, and that both multilayer perceptrons and CNNs can benefit from transfer learning. This study is the first where transfer learning in CNNs aids in the prediction of one experimental property in a small data regime, using data of another experimental property.