A Principal Component Analysis-Integrated Machine Learning Approach for Predicting Gas-Phase VUV/UV Absorption Spectra of Molecular Compounds

This study integrates principal component analysis (PCA) with random forest (RF) modeling to present an enhanced machine learning framework for predicting gas-phase vacuum ultraviolet/ultraviolet (VUV/UV) absorption spectra. This method overcomes a significant drawback of earlier research that directly applied standalone RF models to spectral data, which led to the failure to include notable interwavelength correlations in high-dimensional spectra. A data set of 1397 chemical compounds was used, each represented by 746 spectral absorption intensity values (125-240 nm) and an 86-dimensional feature vector that was generated using molecular structure-based descriptors. PCA was applied to reduce dimensionality while representing 99% of the total variability in the spectral outputs. An RF regressor was then trained using the PCA-transformed outputs. This PCA-integrated method allowed the RF model to compactly capture structured spectral dependencies and lowered training time by more than 15-fold while achieving equivalent or better <i>R</i>² values (up to 0.727) compared to previous work. Moreover, we evaluated a complete 31-feature E-state fingerprint set in comparison to the reduced 16 E-state features used in previous research. The complete set improved accuracy when E-state was used alone but did not consistently enhance multidescriptor models, suggesting partial redundancy. Additionally, weighted PCA showed that key principal components matched with chemically meaningful spectral regions. These results demonstrate the value of PCA not only as a dimensionality reduction technique but also as a method to capture spectral behavior and chemical structure into a more interpretable and scalable predictive model.