DELFI: a computer oracle for recommending density functionals for excited states calculations

Area of Science:

• Computational quantum chemistry
• Materials science
• Machine learning

Background:

• Density functional theory (DFT) is crucial in computational chemistry but selecting appropriate functionals for excited-state calculations (TD-DFT) is challenging.
• Accurate prediction of molecular photophysical properties requires careful consideration of multiple electronic states and transitions.

Purpose of the Study:

• To develop an automated system for recommending DFT functionals for TD-DFT calculations.
• To simplify the selection of functionals for predicting molecular optical properties.

Main Methods:

• Developed a scoring system to assess the accuracy of DFT functionals against high-accuracy references for excited states.
• Created a large database of four million data points evaluating common functionals on organic molecules.
• Trained a graph attention neural network (DELFI) on this data to predict functional performance.

Main Results:

• DELFI provides accurate functional recommendations for TD-DFT calculations, transforming property prediction into a regression task.
• The system was validated by selecting functionals for spiropyran-merocyanine isomers and screening 50,000+ organic photovoltaic molecules.
• An open database and web application were released to facilitate community use.

Conclusions:

• DELFI significantly alleviates the difficulty of choosing DFT functionals for TD-DFT.
• The data-driven approach enables rapid screening and accurate prediction of optical properties for large molecular datasets.