Systematic Assessment of Calibration Strategies in Spectroscopic Analysis: A Case Study of Paracetamol Crystallization

Abstract

Converting spectral data to concentration is beneficial for effective crystallization process monitoring, enabling timely insights into supersaturation profiles. Calibration models are essential in this process, as they transform spectral information into concentration data. While various calibration strategies exist in the literature, they typically involve three stages: Stage 1 for baseline correction, Stage 2 for regressor selection, and Stage 3 for model form selection. In this study, we systematically evaluated all common strategies within each stage, combining them through a Design of Experiments (DoE) approach using a single paracetamol (PCM) and p-acetoxyacetanilide (PAA) crystallization system. The results showed that Savitzky–Golay Second Derivative (SGSD) performed best for baseline correction (Stage 1), while selecting spectral data from a specific range yielded the highest accuracy in regressor selection (Stage 2). For model selection (Stage 3), Partial Least Squares Regression (PLSR), Principal Component Regression (PCR), and Artificial Neural Network (ANN) were assessed with three optimized models deployed to monitor four crystallization runs in real time. During deployment, PLSR demonstrated the most moderate concentration prediction. However, when comparing all three model forms, the standard deviation of predicted concentrations ranged from 4% to 6% for PCM and 10% to 30% for PAA, with similar performance across all models. Validation against offline High-Performance Liquid Chromatography (HPLC) data showed relative errors of 0–12% for PCM, while PAA predictions had higher errors ranging from 0 to 50⁺%, largely due to PAA’s lower concentration range (10–20 g/L) compared to that of PCM (100–350 g/L). These findings indicate that while online models provide useful real-time approximations, precise measurements still require offline validation.