Buy a whistle and some orange cones and lace up your high-tech sneakers. Time to break a sweat.
NIR spectroscopy is a secondary technique. That means that the analyzer isn’t directly measuring water content in pet food kibble or fat in cream cheese. But with a good chemometric software package and some quality reference lab data, you can train it better than a Best in Show German Shepard at Westminster.
BUCHI has already done some heavy-lifting, developing calibrations for key quality parameters across various industries. Check out the BUCHI Application Finder to see if we have a Plug & Play solution already developed for you!
Let’s say you want your at-line NIR to measure protein in dry kibble. First, you make a plan to gather samples from several production batches or across a number of kibble product skews so that the samples you’ve collected have a decent range in protein (and other variables that NIR is sensitive to, like moisture, fat and ash). The rule of thumb for the target range is about 20x the standard error of your reference lab technique. Pour kibble into a sample container, collect the NIR spectra, then send each sample off for analysis by Kjeldahl. Once the Kjeldahl protein measurements come back, you plug that information into your NIR software. Now, each spectrum has a reference property of protein associated with it. The next step is to build a calibration model. It’s the mathematical equation that relates your multivariate X-data (spectra) with univariate Y-data (protein). Luckily for you, the software does all of the heavy lifting, typically using partial least-squares regression, and spits out a calibration equation that you can then use to measure the protein content in future production samples. Splendid.
When you’re looking to do qualitative testing, like 100% inspection of all of your incoming raw materials, the idea is the more or less the same: sample plan, collect NIR data, collect primary data, assign properties, create calibration model using chemometric software. The sampling plan should include a note to gather multiple lots of every raw material (rule of thumb: 5 lots or more). You want to use those lots to train your NIR to “see” and be desensitized to all of the acceptable and expected sources of variation, like vendor, particle size, or moisture content. Once you’ve collected NIR spectra for each lot, ship the samples off to some legit lab that can validate their identity and quality. If the samples pass the test, go back in the software and assign a chemical identity as a property for each spectrum (e.g., “sucrose” or “alanine”). Then, sit back as your chemometric software does something fancy like Soft Independent Modeling of Class Analogy (SIMCA) so that you can use your NIR to test the identity of future incoming samples. This type of analysis also works to establish blend uniformity or finished product conformity. Way quicker than HPLC.
What if the calibration performance takes a hit?
Things could roll along pretty smoothly for awhile and you’ve cut way back on the number of kibble samples you’ve been checking by Kjeldahl for protein. Then something changes; a new sources of variation has entered the fold. Maybe your kibble got an extra boost of fiber in the formulation to keep the terriers regular. Or maybe some new dryer equipment is reducing the moisture content of your kibble lower than when you developed the first calibration model. All of a sudden, your NIR measurements aren’t as accurate as they used to be. Or maybe the analyzer software is spitting out measurements, but they are marked with red flags.
The fact is, formulations evolve, plant equipment ages or is replaced, a record humidity summer sets in. Things change, and when they do, it’s time for a calibration update.
The effort in a calibration update is essentially proportional to the magnitude of change affecting the sample/product/process. If there is a new kibble skew that has slightly higher fiber, add 10-20 sample spectra with Kjehldal reference data to the calibration set, recalculate the model and test the updated model with some new lots. If it works, meaning you’re getting an acceptable standard error of prediction, you’re back off to the races.
If you’re a current BUCHI customer needing support in calibration development or calibration update, contact us!
If this post was enough to wet your whistle, be sure to click the [FOLLOW] button on your browser to get access notifications of future content where we will delve deeper into the details of calibration development, performance and maintenance.