Cost-Effective Accuracy in Molecular Structures via Smart Databases, Topological Features, and Random Forests

We present a cost-effective protocol for predicting accurate equilibrium geometries of organic molecules containing H, C, N, and O atoms. Representative structures are selected from the large QM9 database using chemically grounded descriptors and reoptimized with a composite electronic-structure scheme, yielding a compact yet exhaustive sampling of the reference chemical space. Central to this framework is the concept of atom synthons, continuous topology-aware generalizations of atom types that can be automatically derived from SMILES or Cartesian coordinates and systematically combined into bond-, angle-, and torsion-level features. Building on this representation, we develop a two-tier protocol: (i) feature-based selection ensures broad coverage of bonding motifs with minimal redundancy and (ii) Random Forest regressors provide nonlinear bond-length corrections to low-level geometries, achieving near-spectroscopic accuracy. Validation against both high-level quantum mechanical and semiexperimental structures confirms excellent performance across hydrocarbons, species containing heteroatoms, and sterically congested frameworks. By integrating rigorous validation, interpretable descriptors, and robust nonlinear corrections, our approach bridges the gap between efficient electronic-structure methods and the structural fidelity required for spectroscopy and molecular design. The automated workflow ensures reproducibility and paves the way for applications in thermochemistry, spectroscopy, and force-field parametrization.