Advancing density functional tight-binding method for large organic molecules through equivariant neural networks.

Area of Science:

• Computational Chemistry
• Quantum Mechanics
• Machine Learning

Background:

• Semi-empirical quantum-mechanical (QM) methods offer a balance of efficiency and accuracy for complex molecular systems.
• Parameterization is crucial for QM method reliability and performance enhancement.
• Previous work introduced NN_rep to improve Density Functional Tight-Binding (DFTB3) for small molecules.

Purpose of the Study:

• To introduce the EquiDTB framework, leveraging physics-inspired equivariant neural networks.
• To develop scalable and transferable many-body ΔTB potentials, replacing standard DFTB repulsive potentials.
• To extend ML-corrected DFTB applicability to larger molecules and non-covalent systems beyond training data chemical space.

Main Methods:

• Utilized physics-inspired equivariant neural networks (NN) within the EquiDTB framework.
• Developed many-body ΔTB potentials to parameterize the repulsive part of the DFTB method.
• Applied the framework to compute atomic forces and interaction energies for molecular dimers and explore potential energy surfaces.

Main Results:

• EquiDTB demonstrates enhanced performance over standard Tight-Binding (TB) methods for molecular dimers (S66x8).
• Accurate computation of atomic forces and interaction energies for non-covalent systems.
• Effective exploration of potential energy surfaces for large, flexible drug-like molecules, including isomer transitions and vibrational modes.

Conclusions:

• EquiDTB significantly advances the DFTB method by integrating equivariant neural networks with QM data.
• The framework maintains high computational efficiency while enabling reliable simulations of larger and more complex systems.
• This approach paves the way for accurate and efficient biomolecular simulations.