Chemistry-Informed Machine Learning Framework for Predicting Structural Properties in Osmabenzene Complexes

Metallabenzenes, as representative metalla-aromatic compounds, display a pronounced out-of-plane distortion of the metal center relative to the C5 ring. This deviation from planarity, which contrasts with classical criteria for aromaticity, has been shown to enhance aromatic character and is governed by a σ-orbital antibonding interaction, known as the σ-control mechanism. In this study, we investigate the structural properties of osmabenzene complexes using a chemistry-informed machine-learning approach. A data set of 329 osmabenzene structures was compiled, and mechanistically guided descriptors capturing electronic and steric effects were derived from coordination chemistry and molecular orbital theory. Models incorporating orbital energy descriptors achieved high predictive accuracy and quantitatively confirmed the relevance of the σ-control mechanism. To improve computational efficiency, we further developed a chemically interpretable set of ligand-level descriptors that preserved high model performance (R² = 0.970, RMSE = 1.990°, and MAE = 1.544°) without requiring explicit orbital-level calculations. Shapley additive explanation (SHAP) analysis further identified axial ionization potential and equatorial polarizability as the dominant factors controlling nonplanarity. These findings demonstrate the effectiveness of integrating mechanistic chemical understanding into data-driven modeling and provide a practical framework for predicting and rationally designing functional metalla-aromatic materials.