BUCHI NIRCal - Manual (Page 22)

NIRCal 5.5 Software Manual
22 NIRCal 5.5 Manual, Version A
Min/Max represents the calibration range for quantitative properties. For qualitative properties Min/Max
is shown as 0/1.
1.2.6 Calibration Validation Methods
To be able to judge the performance of a calibration a set of independent validation samples is
necessary.
Validation Set (VS)
Normally all spectra within a project are divided into 2 sets with a suggested ratio of 2/3 to 1/3. The
two sets should be completely independent from each other.

C-Set (Calibration Set)

V-Set (Validation Set)
Spectra in the V-Set are not used for the calibration, the V-Set spectra are used like unknown samples
to judge the quality of the calibration (internal validation set). Only the C-Set spectra are involved in
the loading calculation.
Enough spectra of the sample should be available.
VS can be used for all calibration methods.
Cross Validation (CV)
Cross validation (CV) uses all samples as the calibration set for quantitative calibrations except one
sample (or a small group of samples) which is left out.
Validation is accomplished by predicting the left out samples and by systematically varying the
selection of left out samples. The procedure is time consuming because for each selection a
calibration has to be calculated. The method is especially useful when the total number of samples is
small (< 50 samples).
Full cross validation (FCV) means that n-calibrations are calculated so that one spectrum has been left
out and all other are in a calibration.
Limitations

only available for PCR and PLS;

needs at least 2 CV groups or at least 4 C-Set spectra for one-leave-out (full cross validation;
FCV);

will delete the V-Set spectra selection, in case it is not empty.
1.2.7 Calibration Methods
Qualitative Calibrations / Identification
Target is to identify the membership of a sample to a property group. The property groups can be
chemically completely different or similar to the same substance.
Both implemented method are using PCA:

Cluster Analysis CLU

SIMCA
Quantitative Calibrations
Target is to determine the concentration value such as content in %, OH-value or physical
parameters like density, viscosity.
In NIRCal implemented algorithms are:

Principal component regression PCR

Partial least squares regression PLS

Multiple linear regression MLR
PCR, PLS, PCA (CLU) and SIMCA are principal components based methods, MLR and also library
search with spectra comparison are spectra based methods.