Infrared Spectrum Distortion Phenomenon During the Crystallization Process of L-Carnitine and a Quantitative Study Based on Machine Learning

Attenuated total reflection Fourier transform infrared (ATR-FTIR) is a widely used process analysis technique (PAT) in industrial crystallization, which is generally considered to have no response to solids during crystallization. This work, however, discovered an under-reported distortion in ATR-FTIR spectra during the crystallization of L-carnitine (LC). The observed distorted infrared (IR) spectra appear to exhibit an unexpected blue shift and a large intensity increase of the dissolved L-carnitine (LC) characteristic peak at 1595 cm–1. PXRD validation and probe-wiping experiments attributed such distortion to partial fouling on the probe surface. Traditional spectral quantification based on the external standard method was found to fail when it was applied to distorted IR spectra. To address the quantification challenge, machine learning (ML) models were evaluated for quantifying the solution concentration under fouling conditions. The results revealed that ML approaches, specifically support vector regression (SVR) and shallow neural network (SNN), achieved accurate quantitative analysis of the LC solution concentration. This work not only improves the understanding of ATR-FTIR spectral distortion but also highlights the potential application of ML-integrated PAT in crystallization process monitoring.