FTIR spectroscopy can provide quantitative interpretation, but this is more challenging when spectra lack distinct or diagnostic features, even though the underlying spectral data contain subtle variations that correlate with reaction progress. Machine learning can extract these patterns, but typical workflows require large, experimentally generated datasets, creating a substantial barrier when each dataset must be collected through repeated reaction trials.
To reduce this experimental burden, the authors employed a linear mixture approximation in which each species contributes proportionally to the total absorbance. Because peak intensities scale with molecular abundance, the spectrum of a reaction mixture can be approximated as a weighted sum of the spectra of its individual components. By combining ReactIR spectra of pure reactants and products, synthetic “mimicked spectra” were generated that reproduce the major features of real mixtures. This enables construction of large training datasets using only a limited number of measured spectra.
The Suzuki–Miyaura cross‑coupling reaction was selected as a model system due to its minimal structural changes and correspondingly subtle FTIR signatures. After preprocessing, a neural network trained on mimicked spectra accurately predicted reaction yield from inline ReactIR measurements. The prediction model was capable of detecting subtle differences in reaction conditions that were not evident from physical parameters alone. Integrated into an automated flow platform, the model enabled real‑time yield estimation and efficient closed‑loop optimization. This approach demonstrates that ReactIR, when paired with synthetic spectral generation and machine learning, can serve as a quantitative process analytical technology for data‑efficient reaction development and autonomous optimization.
Ashikari, Y., Tamaki, T., Tomite, K., Yonekura, Y., & Nagaki, A. (2025). Real-time inline-IR-analysis via linear-combination strategy and machine learning for automated reaction optimization. Communications Chemistry, 8(1), 287. https://doi.org/10.1038/s42004-025-01676-y