Applying Machine Learning to Thermal Conductance

Machine learning provides a new approach in materials design. Recent years have seen the use of generative models to algorithmically produce candidate molecules for a variety of tasks. One such model is the variational autoencoder (VAE), which can map discrete information to a continuous, low-dimensional space. We study the ability of these methods to design nanostructures for thermal conductivity: from idealistic harmonic chains to functionalized carbon nanotubes (CNTs). Designing the latter's side-chains is of particular interest for overcoming the severe Kapitza resistance of CNTs and accessing their promising thermal properties. We also highlight the ability of a genetic algorithm (GA) enhanced with a discriminating neural network to optimize molecules for their thermal conductance. Because discriminating GAs are forced to consider a diverse field of molecules, they are an effective way to generate a varied dataset for analysis. It can be shown that molecules sampled in this manner are clustered based on their thermal properties in the latent space learned by the VAE. In addition, studying the GA trajectory may reveal design rules that can narrow the search for good conductors.