Creating a Multivariate PLS-model for estimating ethanol concentrations

Why?

The purpose of measuring ethanol concentrations in this way using a spectrophotometer, was to evaluate ethanol production and in the same time explore some unknown territory. The easiest way to measure ethanol concentrations in our cultures would probably be to use a GC, but this seemed much more fun.

Creating the model

We started out by setting the limits for our model by creating a calibration-set of samples with as diverse properties as possible; ethanol concentration (by adding a known amount), growth phase etc, and at the same time set the procedure for measuring so that all samples are treated in the same way.

We used the Varian Cary 50 broad scan spectrophotometer and scanned near IR from 850nm to 1050nm with a resolution of 1 wavelength and a scan rate of 0.1nm/sec. We chose this part of the spectrum as we had determined that ethanol here has a distinct absorption pattern with low background interference from sugars and other complex contaminants.

Using the spectra from all the samples in the calibration-set we constructed a PLS model using UNSCRAMBLER software. This model can then be used to make predictions on the ethanol concentration in the samples with unknown ethanol concentrations.

Results

Calibration/Validation plot

This plot show the calibration set in blue and the validation set in red. The validation is of the type Cross validation where each sample is taken out of the model and the model is recalculated to fit the remaining samples, and the residual is calculated. This is repeated with each sample and the combined residuals are plotted as the validation curve. In a perfect model the validation sets will be identical, but this is never the case because of various noise. We have in this plot estimate that we have a limit of quantification of 50mM (+-20%) ethanol and a limit of detection of about 12.5mM ethanol.