“…Moreover, potential differences in the spectral process signatures [10] for production-and laboratory-scale samples introduce an additional opportunity for prediction error, especially as the laboratory-scale samples will account for the greatest portion of leverage in the calibration dataset. Furthermore, even a well-designed calibration, having controlled variation in many critical factors, can be expected to require periodic maintenance [8,11,12]. While it is straightforward to design a set of tablets covering all degrees of freedom in the space of composition variation, it can be more difficult to anticipate or control the myriad other factors that may impact tablet NIR spectra.…”