Tree-Based Classification of Functional Groups from IR Spectra: Feature Importance Corroborates Quantum-Chemical Fragment Frequencies

Abstract Fourier-transform infrared (FTIR) spectroscopy is a widely used analytical technique for identifying functional groups (FGs) in chemical compounds, but its interpretation remains labor-intensive and reliant on expert knowledge, especially for complex or multicomponent samples. This study presents a tree-based random forest (RF) machine-learning framework, successfully developed to predict the presence of specific FG from an infrared spectra database. Stratification was employed to address positive class imbalance, maintaining the F1-score drop under 12% across models. All models achieved strong overall performance, with accuracy, F1-score, ROC-AUC, and out-of-bag (OOB) scores exceeding 0.8. Of the 46 FGs modeled, 43 demonstrated acceptable performance, while only three underperformed, as categorized by their Matthews correlation coefficient (MCC) and precision-recall area under the curve (PR-AUC) scores. Threshold tuning further improved model performance, yielding up to a 9% increase in MCC score. Finally, feature importance extracted from the trained models exhibited strong interpretability, correlating with theoretically predicted IR spectra and bridging experimental data with theoretical chemical significance. These results establish a robust, interpretable framework with potential for integration into intelligent monitoring and process optimization systems.