Multivariate spectroscopic calibration is now finding increased indust
rial use in the determination of mixture composition and product quali
ty. Typically, these applications involve measurement of batch process
es or process streams by UV, visible, near-infrared or infrared spectr
oscopy, followed by prediction of product composition or quality with
multiple linear regression or partial least squares calibration models
. Real time predictions of composition or quality measures may then be
used to control the process to increase efficiency, purity, etc. One
obstacle that limits widespread use of this strategy is the lack of ca
libration model ruggedness. Lack of ruggedness in calibration models m
ay manifest itself in the form of large prediction errors following sm
all perturbations in instrument response or slight changes in the samp
le system, fiber-optic probe, or process stream composition. In this p
aper, we describe a strategy for developing rugged calibration models
using artificial neural networks and demonstrate the method on several
NIR process data sets. Fourier transform or principal component prepr
ocessing was used to reduce noise and the number of input measurements
per sample. A large number of neural networks with different network
architectures and random initializations were trained to predict compo
sition using the pre-processed data. A sensitivity analysis was perfor
med with monitoring data sets to screen the resulting networks for one
s that were insensitive to simulated wavelength calibration errors, ba
seline offsets, path length changes or high levels of stray light. Ext
ernal validation data sets were used to demonstrate the ruggedness of
selected neural network calibration models. (C) 1997 Elsevier Science
B.V.