This is part 6 of a tour through PEAXACT - Software for Quantitative Spectroscopy from S-PACT. It demonstrates the Calibration Model and shows how to predict numerical features for unknown samples.
In previous parts we mentioned the three major elements of a model: Pretreatment Model, Spectral Model, and Calibration Model. A Calibration Model converts the outcome of a Spectral Model to quantitative features of interest. PEAXACT provides access to both univariate calibration and PLS calibration (multivariate calibration) from a central dialog, which makes comparison of different methods simple.
- Univariate calibration is used in combination with Integration Models and Hard Models, to convert peak areas or component weights to concentrations or related features.
- PLS calibration (PLS = Projection to Latent Structures or Partial Least-Squares) can be used to directly convert a spectrum into any feature, even mixture properties like e.g. viscosity. (In case you are wondering what the Spectral Model of PLS is: internally it uses a factor model as Spectral Model.)
Remember the Data Table introduced in part 2 of this tour? Those features you provide with your samples can be used as reference values to train and validate the Calibration Model. Calibration results are displayed in a report window where you can inspect and compare the performance of calibration alternatives.
Prediction of Features
Similar to the Integration Analysis or the Component Fitting Analysis, a calibrated model can be used to analyze many unknown spectra, except now the result value is not an area or a component weight, but the actual calibrated feature. This analysis is called "Prediction".
The big advantage of calibration over non-calibrated methods (Integration / Component Fitting) is that predicted values are always accompanied by an uncertainty. Therefore, you should always aim for calibration if you are able to provide reference values, because it is considered best practice to not only predict a value, but also tell how accurate the prediction is.
PEAXACT delivers the Uncertainty for a 95% level of confidence with every predicted value and displays it as an error bar in plots and as a separate column in table reports.
The model is complete now and can either be used for offline analysis or it can be deployed to a field instrument where it automatically predicts features whenever a new spectrum is measured.
Calibration and prediction deal with numerical features. Next we talk about categorical features.
Part 7: Performing Classification