Robust ATR-FTIR Calibration for Concentration Monitoring in Mixed-Solvent Crystallization Processes

In-process monitoring of supersaturation is critical for the efficient development of pharmaceutical crystallization processes. Attenuated Total Reflectance Fourier-Transform Infrared (ATR-FTIR) spectroscopy enables in-line concentration monitoring, but its industrial deployment remains limited by the lack of a standardized and robust methodology for developing calibration models in systems involving multiple solvents. This work evaluates calibration strategies for ATR-FTIR monitoring under varying temperature and solvent composition, with the aim of enabling the rapid development of calibration models while showing that commonly used approaches can remain effective for accurate and transferable predictions. The methodology used in this work incorporates a spectral range, with optional trimming and inclusion of auxiliary process variables, standard normalization, Partial Least Squares Regression (PLSR), and in-run calibration. Its performance was evaluated against manual peak-based linear models and principal component analysis coupled with artificial neural networks (PCA+ANN) across two crystallization systems of increasing spectral complexity: paracetamol in ethanol/water and apremilast in acetone/ethanol. For the paracetamol system, the proposed PLSR approach achieved validation mean absolute errors (MAE) as low as 8.1 g/kg with R2 values above 0.98, representing an order-of-magnitude improvement over manual peak-based models. For apremilast, validation MAEs between 2.5 and 3.3 g/kg were obtained, while maintaining reliable performance during dynamic crystallization experiments. Although PCA+ANN models achieved lower errors on calibration and validation data sets, these gains did not translate into improved robustness during independent crystallization runs, where reduced transferability suggested overfitting. In contrast, the PLSR-based approach consistently provided accurate and reliable concentration predictions across both systems and operating conditions. Therefore, for mixed-solvent crystallization, a carefully implemented PLSR strategy can remain highly effective even when nonlinear models might be expected to offer an advantage, and the added complexity of ANN-based models is not necessarily justified for practical ATR-FTIR calibration in pharmaceutical process development.