Developed by Perten’s own Dr. David Honigs, Honigs Regression takes a fundamentally different approach to NIR calibration – one that grows smarter over time rather than needing to be rebuilt.
Near-infrared spectroscopy is fast. That is one of the reasons food manufacturers value it.
But speed at the measurement stage has always come with a quieter cost: the work of keeping the model accurate as the world around it changes. Products evolve. Crop years shift. New recipe variants appear. Ingredients get reformulated. Production conditions drift. And every time those things happen, someone has to decide whether the current NIR model still fits – and what to do when it does not.
For many manufacturers, that decision leads to the same answer, again and again: build another model. Adjust the bias. Add another channel. Repeat.
This is not a failure of NIR. It is a consequence of the way conventional models are built. Honigs Regression was developed to offer a better way.
Why Calibration Exists in the First Place
NIR spectra are relatively straightforward to collect. The chemistry behind them is not. A calibration bridges the gap: it links spectral data to reference laboratory measurements – moisture, protein, fat, and so on — so that future samples can be predicted quickly, without wet chemistry.
Regression is the mathematical process that builds this link. The challenge is that this relationship is rarely as clean as a straight line, and the real world rarely cooperates with linear assumptions.
The Problem with Global Linear Models
The most widely used method in NIR calibration is Partial Least Squares, or PLS. It is well established and genuinely powerful in many situations. But it works on a fundamentally linear model and in food manufacturing, sample populations are rarely linear or continuous.
Two challenges come up repeatedly.
Non-linearity. Real-world samples rarely follow a neat linear distribution. When the underlying data is non-linear and the model assumes linearity, accuracy suffers – particularly at the edges of the measurement range.
Dumbbell distributions. In production environments, sample sets often cluster at two distinct points with relatively few samples in between. Think of a plant producing two defined product variants. PLS tends to explain the gap between those clusters rather than the variation within them, which is what you actually need to measure accurately. A correlation coefficient can look strong while still being misleading, reflecting the distance between product groups rather than real predictive accuracy within each one.
The practical result is calibrations that need frequent updating, growing numbers of product-specific models, and ongoing bias corrections. This is sometimes called channel creep: a lab accumulates dozens of calibrations measuring essentially the same thing, each requiring its own maintenance. Chasing those biases becomes a standing overhead that never quite goes away.
A Different Starting Point
Honigs Regression, developed by Dr. David Honigs during his time at Perten, begins from a different premise.
Rather than fitting one global model across an entire dataset, it builds a library of previously measured samples and uses that library to predict new ones. When a new sample is scanned, the algorithm identifies the most spectrally similar examples in the library and gives them the most weight in the prediction.
The underlying logic is straightforward: once spectra are properly prepared, samples that look spectrally similar should behave chemically in similar ways. There is no need to force that relationship into a straight line. It is found locally, from the most relevant examples available.
Think of it this way: a global model captures a complex shape with a few large strokes. That is useful, but it smooths over local detail. Honigs Regression builds the same shape from many smaller, locally accurate pieces – preserving structure rather than averaging it away. This makes it particularly well suited to the kinds of sample populations food manufacturing actually produces: clustered, non-linear, and changing over time.
The Role of Pre-Treatments
Before spectral similarity can mean anything, spectra need to be properly prepared. Raw NIR spectra carry more information than chemistry alone. Particle size, scatter, baseline shifts, and instrument-to-instrument differences can all influence the signal in ways that have nothing to do with the analyte being measured.
Pre-treatments such as Standard Normal Variate (SNV), de-trending, and mean centering are designed to remove those unwanted influences without distorting the underlying spectral information. This step matters in any NIR approach, but it is especially important in Honigs Regression, where the prediction depends directly on finding the right reference samples in the library. Good pre-treatment means the model is comparing the right things.
A Model That Learns
The most commercially significant feature of Honigs Regression is this: the library can grow.
As new samples are measured and confirmed with reliable reference values, they can be added directly to the library. The model becomes more representative over time – not by being rebuilt, but by being extended.
For food manufacturers, that changes the maintenance equation considerably.
- New crop year. Add representative samples, confirm with lab values, incorporate into the library. The model now covers that crop year — no recalibration required.
- New product variant. Characterize it properly, add it to the library. The model expands without displacing what it already knows.
- Unusual samples. Rather than discarding edge cases because they disrupt a linear model, they can be retained. That unusual sample improves future predictions when similar material appears again.
In performance comparisons, Honigs Regression has been shown to match Artificial Neural Network (ANN) calibrations across a range of scenarios, and is considerably easier to build and maintain. In one documented example, a calibration built on ruminant animal feed data was progressively extended to include monogastric feed samples simply by adding those samples to the library. With 100 added examples, predictive accuracy on the new material improved substantially – without degrading accuracy on the original dataset. PLS, by contrast, improved on the new material but became less accurate on the original: a trade-off that is unavoidable in a linear system.
One important qualifier: the library only improves when the samples added to it are properly validated. Honigs Regression makes better use of good reference data. It does not eliminate the need for it.
What This Means in Practice
For a food manufacturer, the value is not in the mathematics. It is in the operational outcome.
Fewer separate models to maintain. Less time spent chasing bias corrections every time production conditions shift. More confidence that the NIR system reflects real process variation rather than an idealized snapshot from a previous calibration exercise.
The closer a single calibration can come to accurately predicting a wide range of materials and conditions, the closer NIR moves from being a secondary confirmatory technique to a primary measurement method. One calibration that handles whole wheat, ground wheat, barley, malted barley, flour, and bakery mixes is not a theoretical ideal. It is a direction that Honigs Regression actively supports.
Bias corrections still have a role. But they should act as a diagnostic signal that something has changed and needs attention, not as a standing maintenance task that consumes resources indefinitely.
In Summary
Honigs Regression offers a practical and mathematically grounded alternative to PLS for NIR calibration in food manufacturing. Its key advantages:
- Non-linear modelling that handles real-world data distributions — including dumbbell clusters — without forcing linear assumptions
- Library-based prediction that references the most spectrally similar known samples, rather than fitting a global model
- Continuous learning through library expansion, reducing recalibration effort as production conditions evolve
- Outlier tolerance that allows unusual samples to inform future predictions rather than disrupting the model
To learn more about how Honigs Regression performs across grain, dairy, and snack food applications, contact us to arrange a conversation with one of our NIR specialists.