Similarity Clustering for Representative Sets of Inorganic Solids for Density Functional Testing.

Area of Science:

• Computational materials science
• Quantum chemistry
• Solid-state physics

Background:

• Benchmarking density functional theory (DFT) functionals is complex due to property and material dependencies.
• Existing datasets may contain biases from overrepresentation of similar materials, affecting functional performance evaluation.
• Accurate benchmarking is crucial for developing reliable DFT functionals for materials science applications.

Purpose of the Study:

• To develop a method for identifying chemically distinct solids within DFT benchmarking datasets.
• To propose a strategy for creating smaller, representative datasets that minimize bias.
• To ensure new datasets accurately reproduce the average errors of larger, original sets.

Main Methods:

• Clustering analysis based on the distribution of density gradient and kinetic energy density.
• Development of a rebalancing method to create representative subsets of materials data.
• Application of the method to an existing dataset of 44 inorganic solids.

Main Results:

• Identification of distinct chemical groups within the inorganic solids dataset.
• Generation of a representative subset of seven solids from the original 44.
• Demonstration that the representative set closely reproduces the average errors of the original set.

Conclusions:

• A clustering approach effectively identifies chemically distinct solids for improved DFT benchmarking.
• The proposed method enables the creation of smaller, less biased datasets for reliable functional evaluation.
• Representative datasets are valuable for general benchmarking and training new DFT functionals.