Infrared Spectral Descriptors for Reaction Yield Prediction: Toward Redefining Experimental Spaces

Yield prediction in catalytic reactions is essential for improving chemical process efficiency and product quality. Ligands significantly influence reactivity and selectivity, highlighting the need for descriptors that accurately capture their structural and electronic properties. In this study, we focus on infrared (IR) spectra, which reflects molecular vibrational modes, and propose novel descriptors based on wavenumber information. We evaluated the predictive performance of these descriptors using two datasets: direct Pd-catalyzed arylation and Suzuki-Miyaura coupling reactions. The wavenumber-based IR descriptors outperformed conventional molecular descriptors and structural fingerprints (one-hot encoding, Mordred, MACCS, Morgan fingerprint, RDKit, and density functional theory). Notably, descriptors limited to the fingerprint region (0-1700 cm^-1) effectively captured key molecular features, contributing to both high prediction accuracy and improved chemical interpretability. Our results indicate that IR-based descriptors can achieve strong generalization performance even with small datasets. This approach offers a promising strategy for redefining reaction condition spaces and enhancing the interpretability of predictive models, thereby supporting more informed experimental design.