NN-xTB: density functional accuracy at semi empirical speed with neural network extended tight binding

Area of Science:

• Computational chemistry
• Quantum chemistry
• Machine learning

Background:

• Accurate molecular modeling relies on quantum chemistry methods, but they are computationally expensive for large-scale applications.
• Semi-empirical tight-binding methods offer speed but lack accuracy due to fixed parameters that don't capture local chemical environments.

Purpose of the Study:

• To develop a faster and more accurate molecular modeling approach.
• To enhance semi-empirical tight-binding methods with machine learning for improved chemical adaptivity.

Main Methods:

• Introduced a neural network extension to extended tight-binding (ETB).
• The method learns environment-dependent shifts for Hamiltonian parameters.
• Maintains self-consistency, charge/spin treatment, and long-range behavior.

Main Results:

• Achieved density functional theory (DFT) accuracy at near ETB computational cost.
• Significantly reduced errors in thermochemistry (GMTKN55 benchmark) and vibrational spectra.
• Outperformed leading machine-learned potentials for molecular forces on several datasets.

Conclusions:

• The neural network extension of ETB offers a significant improvement in accuracy and efficiency for molecular modeling.
• This approach bridges the gap between high-accuracy quantum chemistry and fast, approximate methods.
• Enables accurate large-scale and high-throughput molecular simulations across various chemical applications.