AstroSpace v1.1: an exhaustively enumerated CHNO chemical space (≤ 6 heavy atoms) with GFN2-xTB dipole moments and rotational constants for evaluating machine-learning prioritization of interstellar molecules

AstroSpace v1.1: data, code and reference run AstroSpace is an exhaustively enumerated chemical space of 35,482 neutral, closed-shell molecules of carbon, hydrogen, nitrogen and oxygen with up to six heavy atoms. For the 32,458 molecules whose GFN2-xTB calculation converged, it provides: the optimized geometry; the permanent dipole moment and its components along the principal axes of inertia; the rotational constants A, B and C; the total energy. Together with a same-size evaluation protocol, AstroSpace provides a size-controlled benchmark for machine-learning prioritization of interstellar molecules. In this protocol, each confirmed interstellar molecule is ranked only against undetected molecules with the same number of heavy atoms. This record accompanies the manuscript “AstroSpace: A Size-Controlled, Exhaustively Enumerated CHNO Chemical Space for Evaluating Machine-Learning Prioritization of Interstellar Molecules” by A. A. Khairbek, A. Y. A. Alzahrani, S. F. Mahmoud, E. S. Dessoky and R. Thomas (ChemPlusChem, under review). What is new in version 1.1 Dipole-moment components along the principal axes of inertia (|μa|, |μb|, |μc|) for all 32,458 molecules, added to AstroSpace6_xtb.csv. The optimized GFN2-xTB geometries of all 32,458 molecules (AstroSpace6_geometries.xyz). The complete ranked list of undetected molecules (AstroSpace6_ranked_candidates.csv). It comes from a descriptor model trained on every in-range detection reported up to October 2026 (100 positives), and Tables S1 and S2 are updated accordingly. A comparison with the QM9 molecules of the same size range: composition, structural classes, and the same-size AUC of every predictor on both candidate pools (Tables S4–S6, qm9_comparison.json). An extended prospective test. The 21 neutral closed-shell CHNO molecules first reported securely between 2022 and 2026 are ranked by the model trained on the 2021 census only. Their absolute ranks are in Table S3, and the summary counts are in paper_numbers.json. Four tentative or disputed reports are listed separately. A corrected and complete QM9 reference file: all 796 QM9 molecules with at most six heavy atoms, with their QM9 indices, B3LYP/6-31G(2df,p) dipole moments and rotational constants. Pinned software versions (RDKit 2026.03.6, tblite 0.7.0, ASE 3.29.0, XGBoost 3.1.3, scikit-learn 1.9.1) and the executed reference run on Kaggle. The GFN2-xTB dipole moments, rotational constants and energies are identical to those of version 1.0. Files AstroSpace.ipynb: the complete pipeline, covering enumeration, GFN2-xTB properties, evaluation, the QM9 comparison, the prospective test, the candidate list, and all tables and figures. AstroSpace_reference_run_executed.ipynb: the reference run of the manuscript, a single execution on a Kaggle CPU session of about three hours, with all outputs. AstroSpace6_enumerated.csv: the 35,482 enumerated molecules, with canonical SMILES, heavy-atom count, detection labels and QM9 membership. AstroSpace6_xtb.csv: the 32,458 converged molecules, with: the GFN2-xTB dipole moment and its principal-axis components (D); the rotational constants A, B and C (GHz; A is empty for linear molecules); the total energy (eV); the rotational partition function at 10 K and the line-intensity factor μ²/Q_rot; detection labels and QM9 membership. AstroSpace6_geometries.xyz: the optimized geometries (multi-structure XYZ, Å). AstroSpace6_ranked_candidates.csv: every undetected molecule, ranked by the final model. QM9_dipole_reference.csv: the 796 QM9 molecules with at most six heavy atoms. The Supporting Information tables: TableS1_top50_candidates.csv TableS2_top5_per_stratum.csv TableS3_prospective.csv TableS4_qm9_composition.csv TableS5_qm9_structural_classes.csv TableS6_qm9_same_size_auc.csv top50_novel_candidates.csv: the 50 highest-ranked candidates with additional columns, including the Δ-ML-corrected dipole moment with its conformal bounds and both model scores. paper_numbers.json: every number quoted in the manuscript, with the software versions of the reference run. The intermediate results: size_matched_results.json, gin_and_counterfactual.json, dipole_calibration.json, qm9_comparison.json and astrospace_summary.json. README.md and LICENSE (MIT). Detection labelsConfirmed interstellar molecules are taken from the 2021 census (McGuire 2022; astromol v2021.7.0). Of these, 76 have two to six heavy atoms and converged properties, and they form the detected set of the evaluation. Reproducing the resultsRun AstroSpace.ipynb top to bottom on a CPU runtime (Kaggle or Google Colab), with QM9_dipole_reference.csv added as an input. On four cores and without a GPU, it regenerates every number, table and figure of the manuscript in about three hours. The same files are maintained at https://github.com/alikhairbek/AstroSpace.