Computation Aid for Moire Material Electronic Structure Using Targeted Trained Graph Neural Networks

Much attention has been focused on complementing, extending, and sometimes completely replacing ab-initio methods for electronic structure calculations with faster approaches, such as machine learning methods. These can provide more efficient scaling of computational cost, thereby offering insights into structural, thermodynamic and electromagnetic properties in novel larger scale materials, such as twisted bilayer graphene or tunable transition metal dichalcogenides. Results show that in applications related to materials, graph representations with global state attributes in machine learning models deliver more accurate predictions than image or vector based models, as they reflect necessary invariances.

This work focuses on building a graph representation-based machine learning model with global state attributes to predict the electronic structure of uncorrugated twisted bilayer graphene. The model is focused on specialized training, using DFT LCAO to form a database with different twist angles. The calculations are done in steps with special k-points for memory optimization. Predictions are made in competitive times and agree with analytical methods at measured special angles. This leads to an extendable framework that aids in computation of large systems and finding special configurations to study correlated systems