DeePTB is a modeling approach that accurately predicts DFT-like properties of materials using a combination of simplified TB Hamiltonians. In the model, TB Hamiltonians are used based on localized basis functions to accurately represent the electronic structure of various materials. The simplified form of the TB Hamiltonian is parameterized in such a way that it can accurately model the electronic properties of materials with high accuracy. The model also includes a strain-dependent onsite formalism that allows the simulation of strain effects on the material properties.
The neural network architecture of DeePTB involves three main steps: construction of the empirical TB Hamiltonian, extraction of local chemical environments for symmetry-preserving descriptors, and training of the neural network model with ab initio electronic bands as targets. The environment-dependent TB parameters are obtained by correcting the empirical parameters using environment descriptors from the NN model. The detailed NN architecture ensures high expressiveness and accuracy in predicting material properties.
The computational efficiency of DeePTB was demonstrated by comparing it with traditional DFT calculations using different basis sets and XC functionals. The DeePTB model was shown to scale efficiently with system size and had a significant speedup compared to traditional DFT calculations, especially for larger systems. The model’s performance was also evaluated in predicting temperature-dependent properties, optical conductivity, dielectric function, and refractive index with high accuracy.
In addition, the model was successfully applied to predict local density of states and transport properties of materials, such as monolayer graphene. The local DOS and transmission coefficients obtained from DeePTB were compared with DFT-based calculations, showing an excellent agreement between the two methods. The application of DeePTB to simulate quantum transport properties further demonstrated the model’s effectiveness in predicting electronic properties. Overall, the results highlight the accuracy, efficiency, and versatility of the DeePTB approach in various material simulations and electronic property predictions.