Gold-Standard Chemical Database 137 (GSCDB137): A diverse set of accurate energy differences for assessing and developing density functionals

TL;DR

GSCDB137 addresses the need for a gold-standard, diverse benchmark for density functional development by assembling $137$ data sets ($8{,}377$ entries) that span main-group and transition-metal chemistry, noncovalent interactions, and molecular properties including dipoles, polarizabilities, oriented electric-field energies, and vibrational frequencies. It updates and reconciles legacy GMTKN55/MGCDB84 references, prunes spin-contaminated points, and adds new TM- and property-focused data, all with transparent references and supplemental material to enable reproducibility. A broad DFT benchmark across $29$ functionals confirms the general Jacob’s ladder trend, with double hybrids offering the best accuracy but requiring careful handling of basis-set, frozen-core, and spin-symmetry issues; category-specific deviations (notably electric-field and frequency properties) highlight the need to incorporate response quantities into functional design. The database provides practical guidance for practitioners—favoring specific functionals for main-group and TM chemistry, and adopting robust strategies for double hybrids—while serving as a versatile training ground for future non-empirical and ML-based functionals. All data and analysis tools are openly available at the project GitHub repository to support ongoing methodological advancement in computational chemistry.

Abstract

We present GSCDB137, a rigorously curated benchmark library of 137 data sets (8377 entries) covering main-group and transition-metal reaction energies and barrier heights, (intramolecular) non-covalent interactions, dipole moments, polarizabilities, electric-field response energies, and vibrational frequencies. Legacy data from GMTKN55 and MGCDB84 have been updated to today's best reference values; redundant, spin-contaminated, or low-quality points were removed, and many new, property-focused sets were added. Testing 29 popular density functional approximations (DFAs) confirms the expected Jacob's-ladder hierarchy overall but also reveals notable exceptions: functional performance for frequencies and electric-field properties correlates poorly with that for other ground-state energetics. ωB97M-V and ωB97X-V are the most balanced hybrid meta-GGA and hybrid GGA, respectively; B97M-V and revPBE-D4 lead the meta-GGA and GGA classes. Double hybrids lower mean errors by about 25 % versus the best hybrids but demand careful frozen-core, basis-set, and multi-reference treatment. GSCDB137 offers a comprehensive, openly documented platform for stringent DFA validation and for training the next generation of non-empirical and machine-learned functionals.

Figures (5)

• Figure 1: Overview of GSCDB138 Database generation and composition.
• Figure 2: Normalized mean absolute error ratios (NERs) of 29 tested density functionals across seven property categories and their overall mean. Values are relative to the baseline (average of 2nd-4th best hybrid functionals) for each data set.
• Figure 3: Normalized mean absolute error ratios (NERs) for 10 representative functionals on thermochemistry (TC) data sets. Each cell is the NER of a functional on a particular data set, relative to the hybrid baseline (“Standard error” in kcal/mol). For brevity, dispersion-correction labels are omitted.
• Figure 4: Normalized mean absolute error ratios (NERs) for 10 representative functionals on noncovalent interaction (NC) and intramolecular NC (INC) data sets. Each cell is the NER of a functional on a particular data set, relative to the hybrid baseline (“Standard error” in kcal/mol except for O24 and O24x4). For brevity, the names of the dispersion corrections are omitted.
• Figure 5: Normalized mean absolute error ratios (NERs) for 10 representative density functionals across the remaining categories: barrier heights (BH), transition-metal systems (TM), isomerization energy (ISO), electric-field responses (EF), and vibrational frequencies (FREQ). Each cell represents the NER of a functional on a particular data set relative to the hybrid baseline (“Standard error”), expressed in kcal mol$^{-1}$ unless otherwise specified for the corresponding property set or TM sets (see Sec. \ref{['subsec:func_metric']}). For clarity, dispersion-correction labels are omitted.