Accurate Chemistry Collection: Coupled cluster atomization energies for broad chemical space

TL;DR

The paper addresses the need for large-scale, highly accurate thermochemical data across broad chemical space. It introduces the Microsoft Research Accurate Chemistry Collection (MSR-ACC) and its first release MSR-ACC/TAE25, comprising 73,040 TAEs computed at CCSD(T)/CBS via the W1-F12 protocol for closed-shell, neutral, covalently bound molecules up to 5 non-hydrogen atoms from the first three periods. The authors deploy exhaustive graph generation, diverse structure-generation methods, and a careful W1-F12 labeling protocol with filtering based on the %TAE[(T)] diagnostic and singlet–triplet gaps, ensuring reliability for single-reference methods. This dataset provides an unbiased, diverse benchmark for developing and validating data-driven electronic-structure approaches, and it is openly available with canonical train/validation splits to facilitate broad adoption and benchmarking of new methodologies.

Abstract

Accurate thermochemical data with sub-chemical accuracy (i.e., within $\pm$1 kcal mol$^{-1}$ from sufficiently accurate experimental or theoretical reference data) is essential for the development and improvement of computational chemistry methods. Challenging thermochemical properties such as heats of formation and total atomization energies (TAEs) are of particular interest because they rigorously test the ability of computational chemistry methods to accurately describe complex chemical transformations involving multiple bond rearrangements. Yet, existing thermochemical datasets that confidently reach this level of accuracy are limited in either size or scope. Datasets with highly accurate reference values include a small number of data points, and larger datasets provide less accurate data or only cover a narrow portion of the chemical space. The existing datasets are therefore insufficient for developing data-driven methods with predictive accuracy over a large chemical space. The Microsoft Research Accurate Chemistry Collection (MSR-ACC) will address this challenge. Here, it offers the MSR-ACC/TAE25 dataset of 76,879 total atomization energies obtained at the CCSD(T)/CBS level via the W1-F12 thermochemical protocol. The dataset is constructed to exhaustively cover chemical space for all elements up to argon by enumerating and sampling chemical graphs, thus avoiding bias towards any particular subspace of the chemical space (such as drug-like, organic, or experimentally observed molecules). With this first dataset in MSR-ACC, we enable data-driven approaches for developing predictive computational chemistry methods with unprecedented accuracy and scope.

Figures (6)

• Figure 1: End-to-end molecular structure generation. (a) Three methods used for generating new molecular graphs (as SMILES strings). (1) Enumerate all possible graphs, bond types, and atom types before saturating with hydrogen atoms. (2) Sample nonhydrogen atoms, a graph and bond types to satisfy maximum allowed valencies, and saturate with hydrogen atoms. (3) Sample nonhydrogen atoms with explicit hydrogens and then a graph and bond types to satisfy maximum allowed valencies. (b) From molecular graphs to final B3LYP-D3(BJ)/def2-TZVPP optimized structures. Atoms are placed using UFF initially and then optimized using GFN2-xTB which also samples the conformational space. Finally, the geometry is optimized using a meta-GGA (r2SCAN-3c) and a hybrid functional (B3LYP). Throughout, duplicates based on reconstructed molecular graph and matching TAE are merged (indicated by label symbols). Spin states are checked with B3LYP.
• Figure 2: Elemental and structural distributions. (a) Molecular size as atom count. (b) Elemental occurrence. Hydrogen is present in 97.6% of molecules. (c) General 3D shape of the molecules.
• Figure 3: Reconstructed molecular graphs compared across three datasets: MSR-ACC/TAE25, VQM24/DMC, and GDB-9 (also known as QM9). Graphs are reconstructed from 3D structures using a bond model from GFN-FF and a simple heuristic based on a sum of covalent radii. (a) Distribution of element pairs in bonds. (b) Unique 1-st neighbor environments per element. Subsets of MSR-ACC/TAE25 constrained to elemental subspaces are also shown.
• Figure 4: Energetic distributions. (a) Multireference diagnostic %TAE[(T)]. (b) Vertical singlet-triplet gaps S$_0$--T$_1$ from B3LYP. (c) All four components of the W1-F12 total atomization energy. (d) Reference W1-F12 total atomization energy.
• Figure 5: Validation of filtering criteria. (a) Multireference descriptor %TAE[(T)] evaluated on the W4-17 dataset in the small 6-31G(d) basis almost always underestimates the true CBS value (from W4), leading to false negatives but no false positives of multireference character. (b) Singlet--triplet gaps are only overestimated by B3LYP, never underestimated. A random sample of molecules with the B3LYP gap close to zero (inset) has the CCSD(T) gap (from W1w) ranging from 0.0 to 0.4 eV.