Deep Learning Bandgaps of Topologically Doped Graphene

TL;DR

This paper develops deep learning models to accurately predict the bandgaps of doped graphene with various topological dopant arrangements, enabling advanced material design at atomic scales.

Contribution

It introduces a novel CNN-based approach with a material descriptor system for predicting bandgaps in topologically doped graphene, surpassing traditional machine learning methods.

Findings

CNN models achieved R2 > 90% in bandgap prediction.

Deep learning outperformed support vector machine predictions.

Transfer learning improved predictions for larger systems.

Abstract

Manipulation of material properties via precise doping affords enormous tunable phenomena to explore. Recent advance shows that in the atomic and nano scales topological states of dopants play crucial roles in determining their properties. However, such determination is largely unknown due to the incredible size of topological states. Here, we present a case study of developing deep learning algorithms to predict bandgaps of boron-nitrogen pair doped graphene with arbitrary dopant topologies. A material descriptor system that enables to correlate structures with the bandgaps was developed for convolutional neuron networks (CNNs). Bandgaps calculated by the ab initio calculations and the corresponding structures were fed as input datasets to train VGG16 convolutional network, residual convolutional network, and concatenate convolutional network. Then these trained CNNs were used to predict bandgaps of doped graphene with various dopant topologies. All of them afford great prediction accuracy, showing square of the coefficient of correlation (R2) of > 90% and root-mean-square errors of ~ 0.1 eV for the predicted bandgaps. They are much better than those predicted by a shallow machine learning method - support vector machine. The transfer learning was further performed by leveraging data generated from smaller systems to improve the prediction for large systems. Success of this work provides a cornerstone for future investigation of topologically doped graphene and other 2D materials. Moreover, given ubiquitous existence of topologies in materials, this work will stimulate widespread interests in applying deep learning algorithms to topological design of materials crossing atomic, nano-, meso-, and macro- scales.