Chemometrics

In analytical chemistry, instrument calibration is an essential step for most techniques before they can be used to provide meaningful analytical results. In many cases, a given instrument responds not only to the presence of the desired analyte, but also to other constituents in the sample. If the instrument is calibrated using simple standards that only contain the analyte, incorrect results will be obtained when the instrument is used to determine the analyte in real samples that contain the other constituents (known as interferences). In the past, the typical response to this situation would be to develop chemical methods to separate the desired analyte from the interfering sample matrix. Once separated from the sample interferences, the sample would be analyzed with the technique.

While the use of separations to remove interferences is conceptually straightforward, in practice it may be difficult to do quantitatively. Moreover, the separation step increases the complexity of the analytical procedure, a factor that increases the time and cost of the analysis. In process analyses, where sample streams are analyzed "on the fly," separations are generally not possible, and other techniques must be used.

In analytical situations where interferences are present, statistical modeling can often be used to produce a calibration algorithm that gives satisfactory results without the need for chemical separations. To develop this model, a representative set of samples must be obtained and analyzed with existing analytical technology to get reliable values for the analyte. Selection of this calibration set of samples is important to the success of the procedure. In statistical terms, the samples selected for the calibration set must "span the sample space." This means that the samples used in calibrating the instrument should experience all the physical and chemical phenomena that could possibly affect the analytical response with the instrument.

Once a representative calibration set of samples has been selected, instrument response (for example, a spectrum) is obtained for each sample. At this point, various factor-based chemometric techniques such as principal-component regression and partial-least-squares regression can be use to determine what relationships exist between the instrument response (spectral data) and the measured analytical data obtained by the standard method. Model building is certainly not a thoughtless endeavor even though computer programs used to perform chemometric studies may give the impression that the procedure is trivial.

If a satisfactory statistical model can be obtained, its predictive ability must be verified with a second set of samples, known as the validation set. Once again, great care must be used in selecting a representative set of samples for this purpose. Once again, the validation set must be analyzed with the standard method. The goal of this phase of the study is to establish the reliability of the calibration model by comparing the results obtained with the instrument to those obtained by the standard method. If the model can be validated, then the simpler instrumental method can be used in place of the more complex standard method. The hoped for benefits of successful chemometric modeling are increased sample throughput and decreased analytical costs.