Improving the Reliability of, and Confidence in, DFT Functional Benchmarking through Active Learning.

Area of Science:

• Computational Chemistry
• Quantum Chemistry
• Method Development

Background:

• Density functional theory (DFT) calculations require reliable validation of exchange-correlation functionals.
• Current DFT benchmarking data sets are often assembled without a clear strategy, leading to chemical bias and limited transferability.
• Improving the data acquisition process is crucial for enhancing the accuracy and applicability of DFT methods.

Purpose of the Study:

• To develop a data-efficient method for curating benchmarking data sets for DFT functionals.
• To address the limitations of unprincipled data assembly in existing validation approaches.
• To create a more representative and challenging data set for evaluating pericyclic reaction calculations.

Main Methods:

• Employed an active learning strategy to guide the selection of new data points.
• Designed a chemical reaction space by combining reaction templates and substituents around an initial data set (BH9).
• Trained a surrogate model to predict the divergence of activation energies across 20 DFT functionals.

Main Results:

• Identified molecular structures leading to significant DFT functional divergence.
• Demonstrated that the relationship between molecular structure and functional divergence is highly learnable.
• Achieved convergence with fewer than 100 acquired reactions, significantly enhancing the data set.
• Curated a new, more representative pericyclic reaction benchmarking data set.

Conclusions:

• Active learning offers a data-efficient solution for creating robust DFT benchmarking data sets.
• The developed method significantly improves the quality and representativeness of validation data.
• The curated data set reveals substantial changes in DFT functional performance compared to the original subset.