문서 공유
How pre-calibrations assist quick implementation of near-infrared spectroscopy
Wouldn’t it be nice to start your analysis at the push of a button? Insert the sample, close the lid, and start the measurement – this is all it takes with NIR spectroscopy pre-calibrations.
This is part four in our series about NIR spectroscopy. In this installment, we outline in which cases NIR spectroscopy can be implemented directly in your laboratory without the need for any method development. This means that for these applications your instrument is immediately operational to deliver accurate results – right from day one.
The following topics will be covered (click to jump to the topic):
The advantage of pre-calibrations
In our last installment (How to implement NIR spectroscopy in your laboratory workflow), we showed how you can integrate a newly received NIR spectrometer in your laboratory workflow with a real application example. This process is depicted in Figure 1.
Most of the work consists of creating a calibration set. You have to measure approximately 40–50 samples across the expected parameter range with a primary method. Using a NIR software, you have to link the resulting values to the NIR spectra that were recorded for the same samples (Figure 1: Step 1).
Thereafter, a prediction model needs to be created by visually identifying the spectral changes and correlating these changes to the values obtained from the primary method (Figure 1: Step 2). After validation by the software, a prediction model is available for use in routine measurements.
The process described above requires some effort and is of significant duration because, in many cases, the samples spanning the concentration range first need to be produced and collected. Therefore, it would be very beneficial if steps 1 and 2 could be omitted so that the NIR instrument can be used immediately from day one.
This is not just wishful thinking, but rather the reality for specific applications with the use of pre-calibrations.
What are pre-calibrations?
Pre-calibrations in NIR spectroscopy are prediction models that can be used immediately and provide satisfying results right from the beginning. These models are based on a large number of real product spectra (between 100–600) covering a wide parameter range.
This means that calibration set creation and prediction model creation and validation (Figure 1: Steps 1 and 2) are not required. Instead, the pre-calibration prediction model can be used directly for routine analysis of unknown samples, as illustrated in Figure 2.
How do pre-calibrations work?
Each pre-calibration comes as a digital file that must be imported into the NIR software, such as Metrohm Vision Air software.
- Install the new NIR instrument (including the Vision Air software).
- Create a method containing measurement-specific settings, such as measurement temperature and sample vessel type used.
- Import the pre-calibration and link it to the method.
That’s all that is needed!
The instrument is now ready to deliver reliable results for routine measurements. It is advised to measure a few control samples of known values to confirm that the pre-calibration provides acceptable results.
Optimizing the pre-calibration
In some cases, the results obtained on control samples with the pre-calibration are not completely acceptable. There can be various reasons for this and in general, three different cases are distinguished:
- The results obtained with the control samples deviate only slightly from the expected values.
- The results are acceptable, but the standard error is somewhat on the larger side.
- The results deviate significantly.
We will go through each of these cases below and provide recommendations.
Case 1: The results obtained with the control samples deviate only slightly from the expected values
If the value obtained from the control samples deviates only slightly, a slope-bias correction is the recommended solution. The process is illustrated in Figure 3.
In the top diagram, you see that the values from the pre-calibration deviate consistently over the whole range. In this situation, it is possible to perform a slope-bias correction on the measured model in the Vision Air software. After this has been done, the results fit very well (Figure 3 – bottom).
Case 2: The results are acceptable, but error is somewhat on the larger side
In most cases, this behavior is observed if the range of the pre-calibration is much larger than the range that the analyst is interested in.
Consider for example, measurement of a value at the lower end of the overall range. The error of the pre-calibration is calculated over the entire range, and therefore the impact of the average error (SECV = standard error of cross validation) is much larger on values on the lower end compared to values in the middle of the complete range. This is exemplified in Figure 4 and Table 1.
Table 1. Figures of merit for the different regions of the pre-calibration from Figure 4. Note the much smaller SECV for the range 0–36 compared to the SECV for the full range of 0–200.
| Range | R2 | SEC | SECV |
| 0–200 | 0.996 | 3.8 | 3.9 |
| 0–36 | 0.994 | 0.77 | 0.81 |
| 32–109 | 0.986 | 3.3 | 3.8 |
| 91–200 | 0.977 | 3.6 | 3.7 |
The recommended action in this case is to remove certain ranges of the pre-calibration, leaving in only the range of interest.
From Table 1, it is clear that the SECV for the whole range (0–200) is much higher than the SECV of the smaller range (0–36). This means that when removing the samples corresponding to the higher ranges from the pre-calibration (leaving only the range of 0–36 in), the resulting modified pre-calibration gives a lower SECV.
Case 3: The results deviate significantly
There can be several reasons behind an unsatisfying prediction performance. We will discuss two reasons.
In the first example, consider the possibility that the samples provided for analysis are proprietary. For instance, certain manufacturers produce unique, patented polyols. These proprietary substances are not included among the standard collection of sample spectra in the pre-calibration. Thus, the pre-calibration does not provide acceptable results for such proprietary samples.
Another example is shown in Figure 5. It can be observed that the values from the primary method (blue data points) deviate significantly from the values obtained from the pre-calibration model.
This example is taken from a real customer case which we have observed. At first, we were a bit puzzled when checking the measurement results, but the reason became clear after speaking with our customer. They had chosen to measure the primary values (hydroxyl number) via manual titration and not, as recommended, with an automatic titrator from Metrohm.
Therefore, the reason for the unsatisfying performance is the poor accuracy of manual titration of the control samples and has nothing to do with the quality of the pre-calibration.
For more information about manual titration, read our blog post below about the main error sources:
NIR spectroscopy pre-calibrations from Metrohm
Metrohm offers a selection of pre-calibrations for a diverse collection of applications. These are listed in Table 2 together with the most important parameters of the pre-calibration. Click on the links to get more information.
Table 2. Overview of available pre-calibrations for the Metrohm Vision Air software.
| Pre-calibration | Selected Important Parameters |
| Polyols | Hydroxyl number (ASTM D6342) |
| Gasoline | RON, MON, anti-knock index, aromatics, benzene, olefins |
| Diesel | Cetane index, density, flash point |
| Jet Fuel | Cetane, index, density, aromatics |
| Palm oil | Iodine value, free fatty acids, moisture |
| Pulp and Paper | Kappa number, density, strength parameters |
| Polyethylene (PE) | Density, intrinsic viscosity |
| Polypropylene (PP) | Melt Flow Rate |
| Polyethylene Terephthalate (PET) | Intrinsic viscosity, acid number, and others |
| Polyamide (PA 6) | Intrinsic viscosity, NH2 and COOH end groups |
| Cannabis | THC, CBD, and CBG content; moisture |
| Stool analysis | Fat, calorie, and nitrogen content; moisture |
Conclusion
Pre-calibrations are prediction models based on a large number of real product spectra. These allow users to skip the initial model development part and make it possible to use the instrument from day one, saving both time and money.
