Optimization of the memristor fabrication based on graph convolutional network

Abstract

In this article, a machine learning model for accurately predicting the performance of unknown memristors is constructed by employing a graph convolutional network approach. Thickness and elemental composition are used to transform memristors into graph-structured data. This model exhibits high accuracy and, based on extensive training with a certain type of memristor data, can be applied to novel memristors and give rapid predictions of the performance with only a small-batch sample reported in the literature, showing the potential for excellent transfer learning. This model is also applied to predict the performance of halide memristors, which have received less attention in current research, and it is indeed that a halide perovskite memristor with potential high switching ratio is predicted.