Abstract
Machine learning (ML) is a methodology for uncovering patterns and principles hidden within empirical measurement data. Among these, generative modeling using a variational autoencoder (VAE) is a powerful technique for generating new data based on learned ML model, and it has demonstrated success in accurately generating molecular structures. If membrane permeation property can be evaluated from molecular structures in a non–empirical manner, it would enable the design of novel, high–performance separation membranes. This paper focuses on providing an overview of applications of the VAE generative model, along with the characteristics and utilization of the latent space. As a concrete example of its application, research on the molecular structure design of carbon dioxide separation membranes was presented.
