DeePaTB: A deep learning powerd semi-empirical quantum mechanical method

Deep learning approaches achieve high accuracy with low computational cost by leveraging feature representations from modern low-level quantum methods. We present Deep Atomic Density-Based Tight-Binding (DeePaTB), a novel machine learning-based semi-empirical quantum mechanical (ML-SQM) framework, developed upon our recently proposed atomic density-based tight-binding (aTB) method, which can generate the descriptor by Amesp, "eigenvalue of the local density matrix." This neural network-enhanced semi-empirical quantum mechanical model demonstrates remarkable computational efficiency and transferability across diverse chemical systems. The framework successfully models closed-shell configurations, open-shell configurations with unpaired electrons, and environments with external electric fields. Through reaction data training, DeePaTB achieves density functional theory (DFT)-level accuracy while maintaining the computational efficiency characteristic of semi-empirical quantum mechanical methods. This work demonstrates a versatile framework that bridges the accuracy-efficiency trade-off in quantum chemical calculations, offering a promising tool for large-scale computations and reaction mechanism studies.