Abstract
Accurate thermochemical data with sub-chemical accuracy (within 1 kcal mol$^{-1}$ of the empirical ground truth) are essential for advancing computational chemistry methods. However, existing datasets that reach this level of accuracy remain limited in size or scope. This hinders the development of data-driven methods with predictive accuracy across the broad chemical space of closed-shell, neutral molecules. Here we present Microsoft Research Accurate Chemistry Collection (MSR-ACC) and its first release, MSR-ACC/TAE25, comprising 73,040 total atomization energies at the CCSD(T)/CBS level obtained with the W1-F12 thermochemical protocol. The dataset is constructed to exhaustively cover the chemical space of closed-shell, charge-neutral, covalently bound equilibrium molecular structures containing up to 5 non-hydrogen atoms drawn from elements up to argon and lacking significant multireference character. The dataset and its canonical train and validation splits are openly available on Zenodo in the QCSchema format under the CDLA Permissive 2.0 license. This first release of MSR-ACC enables data-driven approaches for developing predictive computational chemistry methods with unprecedented accuracy and scope.
Introduction
In computational chemistry, achieving "chemical accuracy"—predicting molecular properties within 1 kcal/mol of experimental values—is considered the gold standard. While high-level methods like CCSD(T) can reach this threshold, their extreme computational cost limits their use to small systems. Conversely, faster methods such as Density Functional Theory (DFT) or machine learning often lack the necessary precision for high-stakes chemical modeling.
A significant bottleneck in developing more accurate, efficient models is the lack of high-fidelity data. Existing datasets are often small or biased toward organic "drug-like" molecules. Furthermore, many focus on properties where systematic errors might cancel out. This research introduces the Microsoft Research Accurate Chemistry Collection (MSR-ACC), specifically focusing on Total Atomization Energy (TAE). TAE represents the energy required to break a molecule into its constituent atoms, serving as a rigorous stress test for theoretical methods because it offers no systematic error cancellation, providing a true upper bound on a method's error.
Methodology: The Data Generation Pipeline
The MSR-ACC/TAE25 dataset was constructed using a multi-stage pipeline designed to maximize chemical diversity and ensure sub-chemical accuracy.
The process begins with molecular graph generation using three distinct strategies: brute-force combinatorial enumeration for small systems, degree sequence sampling, and an autoregressive GPT-2 transformer trained to generate novel, valid molecular structures. These graphs are then converted into 3D structures and refined through a cascade of increasingly accurate computational methods, culminating in B3LYP-D3(BJ)/def2-TZVPP optimizations.
Rigorous filtering was applied to exclude unstable structures (saddle points), triplet ground states, or molecules with significant multireference character where standard coupled-cluster methods might fail. The final "gold standard" labels were calculated using the W1-F12 thermochemical protocol, which approximates the CCSD(T)/CBS limit with a mean absolute deviation of only 0.51 kcal/mol compared to the more expensive W4 theory.
Dataset Characteristics and Diversity
The initial release, MSR-ACC/TAE25, contains 73,040 molecules. A key feature of this collection is its broad chemical scope. Unlike datasets like GDB-9, which are predominantly organic, MSR-ACC is 54.9% inorganic. It includes elements from H to Cl (Li, Be, B, C, N, O, F, Na, Mg, Al, Si, P, S, Cl) in diverse bonding environments, including challenging s- and p-block configurations.
The use of AI-driven discovery via GPT-2 allowed for the exploration of chemical spaces that traditional rule-based enumerators often miss. This diversity ensures that models trained on MSR-ACC are exposed to a wide variety of linear, planar, and complex 3D molecular geometries.
Results and Impact
The MSR-ACC/TAE25 dataset represents a massive leap in the availability of high-quality thermochemical data, increasing the scale of "gold standard" labels from hundreds or thousands to over 70,000. By benchmarking 12 common DFT functionals against this data, researchers showed that DFT errors follow a normal distribution, facilitating the identification of "outlier" molecules where standard approximations break down.
By providing a true upper bound on method errors through TAEs, this collection enables the development of next-generation machine learning models and the refinement of DFT methods with unprecedented predictive accuracy across a broad chemical space. The entire dataset is available on Zenodo in the QCSchema format, supporting the broader research community's efforts in data-driven computational chemistry.
Limitations and Future Work
Currently, the dataset is limited to closed-shell, neutral molecules with up to five non-hydrogen atoms due to the extreme computational cost of the W1-F12 protocol. Future expansions aim to include larger molecules, radicals, ions, and transition states. Additionally, while non-relativistic electronic structures are the current focus, future iterations may incorporate scalar relativistic and spin-orbit corrections for even higher precision.
Summary updated