Industry NIR applications would require the right sampling

As an NIR application specialist, when faced with the challenge of developing a new, unknown application, I tend to look for comparable cases that have been shown to work (or not work) in the past.

In theory, if we simplify the theory a little, NIR works great for determining anything that’s made from organic matter and that has molecular bonds that include C, H, N, S, etc, because these elements tends to form dipolar bonds on a molecular level, which is what we detect with NIR. And it’s not supposed to be able to predict stuff that does not form dipolar bonds, such as metals and minerals. Still, thousands of NIR instruments are used daily for the determination of such parameters, at times with very high precision, such as for ash in wheat flour. Knowing this, I am careful in making general assumptions about what can and can’t be done with NIR as it’s a very empirical method.

For these “exotic” parameters, we are generally looking at some sort of secondary correlation that can be hard to identify and understand. Thus, great care must be taken when validating calibrations for these parameters, as they may yield unexpected results when conditions change. We can take ash in poultry feed as an example. It is pretty straightforward to develop a calibration for ash – as long as you don’t add ingredients such as calcium carbonate, which is basically a mineral. I can interpret this as long as we have a mixture of vegetable and animal based ingredients. The minerals (P, K, Ca, etc) are basically integrated in organic structures that can be identified with NIR. But if you add inorganic sources, these correlations are broken.

So, when looking at the questions presented, I would compare it to the task of measuring amino acids with NIR. This is widely done for raw materials, where amino acids are easier to identify than they are in mixed feeds. It also makes more sense from a feed formulation point of view, as the potential savings are realized when you define how to mix your raw materials and additives. Once they are processed, all you can do is verify that you got what you aimed for.

Now, applying that concept to Phytic acid, I would assume that it can be determined in raw materials. It has indeed been shown that Phytic acid can be determined in raw materials with high precision  and in finished feed at a level that can be used for classification.

However, these are scientific studies done on a limited number of samples under lab conditions. When implementing these kinds of parameters in the industry, great care must be taken and the calibrations need to be evaluated on a regular basis. Samples from new regions or unusual process conditions could give unexpected results. That makes the task unfeasible for individual feed mills. Larger groups or specialized providers that have access to a wide variety of samples can implement a program where they collect samples from participating feed mills on a regular basis for evaluation. This would enable them react to changing conditions quickly by publishing calibration updates.


Dan Shiley
Dan Shiley

Excellent consideration David! It is very important to understand whether you are working with a direct or indirect relationship. Having a pure component spectrum of your constituent can help tremendously.  The model factor loadings plot should have features in the same position as your pure component spectrum that has been treated the same as the model is using.  In other words, if the model uses a Sav-Gol derivative the Sav-Gol derivatized spectrum should look similar.  If it does, you have a direct relationship, if not it is an indirect one and everything David says should be carefully considered.  It is also critically important to establish an ongoing calibration validation plan. This should include the frequency of follow-up testing on validation samples.  That way if the relationship is indirect you will be able to quickly discover when this relationship breaks in the case of adding inorganic mineral sources.