Accelerate the design of new superhard carbon allotropes in Pca21 space group: High-throughput screening and machine learning strategies

Abstract

As an efficient calculation and screening method, high-throughput can discover and optimize new materials, shorten the development cycle and cost of new materials. However, using high throughput for material screening and computation, a large amount of computing resources and storage space are indispensable. To accelerate the design of novel superhard carbon materials, we combined machine learning methods with high-throughput computing to construct three machine learning models: support vector machine regression, random forests, and artificial neural networks, and mined data from existing material databases to select 1276 structures as datasets for the model to predict the volume modulus and shear modulus. Through comparative analysis, the optimal model was selected to predict the bulk modulus and shear modulus of the structures obtained by high-throughput calculations, and the prediction results of the model were verified by density functional theory (DFT) calculations, and 8 superhard carbon allotropes in the Pca21 space group were eventually found.