Molecular Systems

GMTKN55

Summary

Performance in evaluating gas-phase chemical accuracy for main group thermochemistry, kinetics and noncovalent interactions, through 55 test sets, totalling 1,505 relative energies, categorised into five chemical domains.

Metrics

  1. Small systems

Weighted mean absolute deviation (MAD) of basic properties, such as atomic energies, ionisation potentials, and electron affinities, of small systems.

For each system, the relative energy is calculated and compared to the reference energy. The MAD is calculated for each of the subsets within this category, including only neutral singlet systems. A weighted sum is calculated by multiplying each subset error by a weight and the number of systems in the subset. This is divided by the total number of systems within these subsets.

  1. Large systems

Weighted mean absolute deviation (MAD) of reaction energies of large systems and isomerisation energies.

Same as (1), for the appropriately categorised subsets.

  1. Barrier heights

Weighted mean absolute deviation (MAD) of reaction barrier heights for transition state energetics for fundamental organic reactions.

Same as (1), for the appropriately categorised subsets.

  1. Intramolecular NCIs

Weighted mean absolute deviation (MAD) of intramolecular noncovalent interactions for conformational energetics and hydrogen bonding.

Same as (1), for the appropriately categorised subsets.

  1. Intermolecular NCIs

Weighted mean absolute deviation (MAD) of intermolecular noncovalent interactions for dimers, clusters, and host-guest complexes.

  1. All (WTMAD)

Weighted mean absolute deviation (MAD) of all subsets.

Same as (1), for all subsets.

Computational cost

Low: tests are likely to take minutes to run on CPU.

Data availability

Input structures:

  • L. Goerigk, A. Hansen, C. Bauer, S. Ehrlich, A. Najibi, and S. Grimme, A look at the density functional theory zoo with the advanced gmtkn55 database for general main group thermochemistry, kinetics and noncovalent interactions, Physical Chemistry Chemical Physics 19, 32184 (2017).

Reference data:

  • Same as input data

  • CCSD(T)

Wiggle150

Summary

Performance in predicting relative energies between 150 strained conformers of adenosine, benzylpenicillin, and efavirenz molecules (50 each) and their geometry optimised structures.

Metrics

  1. Relative energy MAE

Accuracy of relative energy predictions.

For each molecule, 50 relative energies are calculated by comparing the predicted energy of the DLPNO-CCSD(T)/CBS geometry optimised structure to the energies of its 50 strained conformers. The mean absolute error is reported over all 150 conformers.

Computational cost

Low: tests are likely to take less than a minute to run on CPU.

Data availability

Input structures:

  • Brew, R. R. et al. Wiggle150: Benchmarking Density Functionals and Neural Network Potentials on Highly Strained Conformers. J. Chem. Theory Comput. 21, 3922-3929 (2025).

Reference data:

  • Same as input data

  • DLPNO-CCSD(T)/CBS

BMIM Cl RDF

Summary

Tests whether MLIPs incorrectly predict covalent bond formation between chloride anions (Cl⁻) and carbon atoms in 1-butyl-3-methylimidazolium (BMIM⁺) cations. Such Cl-C bonds should NOT form in the ionic liquid under normal conditions.

This benchmark runs NVT molecular dynamics simulations of BMIM Cl at 353.15 K and analyses the Cl-C RDF to detect any unphysical bond formation.

Metrics

  1. Cl-C Bonds Formed

Binary metric indicating whether unphysical Cl-C bonds formed during the MD simulation.

The Cl-C RDF is computed from the MD trajectory. If the RDF shows a peak (g(r) > 0.1) at distances below 2.5 Å, this indicates bond formation and the model fails the test.

  • 0 = no bonds formed (correct physical behaviour)

  • 1 = bonds formed (unphysical, model failure)

Computational cost

Medium: tests require running 10,000 steps of Langevin MD for a system of 10 ion pairs, which may take tens of minutes on GPU.

Data availability

Input structures:

  • Generated using molify from SMILES representations of BMIM⁺ (CCCCN1C=C[N+](=C1)C) and Cl⁻ ions, packed to experimental density of 1052 kg/m³ at 353.15 K.

  • Zills, F. molify: Molecular Structure Interface. Journal of Open Source Software 10, 8829 (2025). https://doi.org/10.21105/joss.08829

  • Density from: Yang, F., Wang, D., Wang, X. & Liu, Z. Volumetric Properties of Binary and Ternary Mixtures of Bis(2-hydroxyethyl)ammonium Acetate with Methanol, N,N-Dimethylformamide, and Water at Several Temperatures. J. Chem. Eng. Data 62, 3958-3966 (2017). https://doi.org/10.1021/acs.jced.7b00654