Data-driven approach for benchmarking DFTB-approximate excited state methods

Area of Science:

• Computational Chemistry
• Quantum Chemistry
• Materials Science

Background:

• Approximate density-functional tight-binding (DFTB) methods are widely used for excited state (ES) calculations.
• Existing DFTB ES methods have limitations due to approximations to density-functional theory (DFT).
• Benchmarking these methods is crucial for understanding their accuracy and applicability.

Purpose of the Study:

• To chemically benchmark approximate DFTB excited state (ES) methods within the DFTB+ suite.
• To identify limitations of these methods by comparing them against more accurate reference data.
• To provide recommendations for improving DFTB ES calculations.

Main Methods:

• A chemically-informed, data-driven approach was employed.
• The QM8 machine learning dataset, containing low-detail ES data, was utilized.
• First singlet-singlet vertical excitation energies (E1) from approximate DFTB methods were compared to those from coupled cluster methods (CC2).

Main Results:

• Clear trends in E1 prediction error distributions were identified across nearly 21,800 organic molecules (GDB-8 chemical space).
• The accuracy of DFTB ES methods showed a strong dependence on the chemical identity of the molecules.
• Valuable insights into the approximations made in DFTB ES methods were extracted.

Conclusions:

• The study highlights the strengths and weaknesses of current approximate DFTB ES methods.
• Recommendations for overcoming limitations and improving the accuracy of DFTB ES calculations are provided.
• The findings contribute to the reliable application of DFTB methods in computational chemistry.