Statistically representative databases for density functional theory via data science.

Area of Science:

• Computational Chemistry
• Electronic Structure Theory

Background:

• The proliferation of data for density functional theory (DFT) methods necessitates efficient assessment and parameterization strategies.
• Existing databases for DFT calculations are extensive, posing significant computational challenges.

Purpose of the Study:

• To introduce a novel cluster analysis technique for creating statistically significant DFT databases with a manageable number of data points.
• To develop a new database (ASCDB) that retains comprehensive chemical information from a large dataset while drastically reducing computational cost.

Main Methods:

• Applied a novel cluster analysis technique to a dataset of nearly 350,000 data points without a priori chemical assumptions.
• Generated a reduced database (ASCDB) comprising only 200 data points.
• Classified chemical properties into non-covalent interactions, thermochemistry, non-local effects, and unbiased calculations.

Main Results:

• The ASCDB database, containing 200 data points, effectively captures the chemical information of a much larger dataset.
• Computational cost is reduced by several orders of magnitude compared to using the full dataset.
• ASCDB demonstrates comparable performance to larger databases like GMTKN55, MGCDB84, and Minnesota 2015B for evaluating DFT methods and basis sets.

Conclusions:

• The novel cluster analysis technique successfully generates a compact yet informative database (ASCDB) for DFT calculations.
• ASCDB serves as an efficient and inexpensive tool for the initial evaluation and parameterization of electronic structure theory methods.
• The findings suggest a paradigm shift towards smaller, data-rich databases for computational chemistry research.