A new procedure for calibrating multiple instruments is presented in which spectra from each are used simultaneously during the construction of multivariate calibration models. The application of partial least-squares (PLS) and genetic regression (GR) to the problem of generating these hybrid calibrations is presented. Spectra of ternary mixtures of methylene chloride, ethyl acetate, and methanol were collected on a dispersive and a Fourier transform spectrometer. Calibration models were generated by using differing numbers of spectra from each instrument simultaneously in the calibration and prediction sets, and then validated by using a set of spectra from each instrument separately. Calibration models were found that perform well on both instruments, even when only a single spectrum from the second instrument was used during the calibration process. As a benchmark, comparison with PLS showed that GR is more effective than PLS in building these hybrid calibration models.

The ability of genetic regression (GR) to correct for wavelength drift in instrument responses was investigated in single- and multiinstrument calibrations. Sample spectra of ternary mixtures were collected on two near-infrared (NIR) spectrometers, one a dispersive instrument and one a Fourier transform instrument, with different resolutions. In the first and second cases, calibration models were generated with the use of spectra collected on a single instrument. For the third case, hybrid calibration models (HCMs) were built in order to combine spectra from two instruments into one calibration model. In order to simulate a wavelength shift, some of the spectra were shifted along the wavelength axis from 0 nm to 6 nm. The performance of GR was poor when calibration models produced from unshifted spectra were used to predict shifted spectra. However, the inclusion of shifted spectra in the calibration model corrected the prediction of shifted and unshifted spectra at levels similar to those of models built and evaluated by using only unshifted spectra.

Genetic regression (GR) is an application of genetic algorithms to the problem of producing optimal calibration models by wavelength selection. GR has been shown to provide excellent calibration models under many conditions that typically result in poor calibration models with the use of other multivariate techniques. In this study, GR is applied to the calibration of the components of a ternary mixture with the use of near-infrared spectroscopic data. To determine how close GR comes to the true global optimum, a random search of the possible solutions was performed and the distribution of the solutions' predictive abilities determined. Through this study it has been determined that GR is capable of searching through extremely large search spaces and eliminating over 99.9999% of the unsuitable solutions in a matter of minutes. GR is also capable of finding multiple solutions of similar quality, something not available in many other calibration techniques. Comparison with results from partial least-squares (PLS) is also included.