spectroscience.com / blog / instrument path / part 4
Instrument path · Part 4 of 4Move a calibration

NIR calibration transfer: why a second instrument disagrees and how to fix it

Vitaly Kirkpatrick · first published November 24, 2025 · updated November 2026 · 11 min read

Short answer: a calibration built on one NIR instrument rarely gives identical numbers on another, even the same model, because each unit sees the spectrum slightly differently. Calibration transfer measures a small set of samples on both instruments, corrects the difference with a bias and slope adjustment or a spectral standardization, and then proves the result on samples neither instrument was tuned on. When the differences are too large, rebuild instead.

Why does a calibration give different results on a second NIR instrument?

A calibration is a set of weights on specific wavelengths. If the second instrument places those wavelengths a fraction of a nanometre differently, or records them with a different bandwidth or detector response, the same weights produce a different number. The sample has not changed; the way it is seen has.

The usual sources of spectrometer variability, roughly in the order I check them:

  • Wavelength axis. A small shift moves every band. It matters most where bands are steep.
  • Bandwidth and resolution. A broader slit or a different design blurs bands differently.
  • Detector response and stray light. These change the absorbance scale, especially at the ends of the range.
  • Sample presentation. A different cup, window, spinner or probe path changes how light meets the sample.
  • Temperature and environment. A warm plant floor and a cool lab shift water bands and baselines.
Same sample, two instrumentsInteractive · simulated spectra
The cyan curve is the master instrument; the violet curve is the same sample on the second instrument. Move the sliders to add a wavelength shift or a scale difference and watch the predicted protein move, then apply standardization, which maps the second instrument's spectrum back onto the master's. Simulated in your browser; the numbers are illustrative.
Two identical compact dark NIR analyzers side by side on a bench, each with a cup of ground grain on its sample window
Two units of the same model still see a sample slightly differently. Transfer starts by scanning the same samples on both. Illustration, not a specific product.

How do you troubleshoot spectrometer variability before a transfer?

Fix what can be fixed in hardware and handling first. A transfer method can hide a problem, but it cannot cure a dirty window or a different sample cup.

  • 1Check both instruments against their standardsRun the wavelength and photometric checks your vendor specifies on both units, after warm-up, and fix anything out of tolerance.hardware
  • 2Match the presentationSame cup type, same fill and packing method, same grind, same sample temperature on both instruments.handling
  • 3Scan the same samples on bothUse the same physical samples, scanned close in time, so the samples themselves cannot differ.data
  • 4Compare the spectra, not just the numbersSubtract the mean spectra. A shifted wavelength axis shows up as a wiggle around every band; a scale difference shows up as a slope or offset.diagnose
  • 5Look at the environmentIf one instrument sits on a warm production floor, compare spectra at matched temperatures before blaming the optics.environment
Key point

Diagnose before you correct. The shape of the spectral difference tells you which transfer method can work, and whether any of them should.

Which calibration transfer method should you use?

MethodWhat it correctsWhat it needsBest when
Bias / slope correctionan offset or tilt in the predictionsreference values for a set of samples scanned on the new instrumentinstruments are very similar and the error is a constant offset
Direct standardization (DS)the whole spectrum of the new instrument mapped onto the masterthe same transfer samples scanned on bothdifferences are broad and smooth
Piecewise direct standardization (PDS)local differences, wavelength by wavelength, including small axis shiftsthe same transfer samples scanned on bothsmall wavelength shifts and bandwidth differences
Robust pre-treatment or global calibrationmakes the model less sensitive to instrument differences in the first placecalibration data from several instruments or conditionsmany instruments or sites share one model
Rebuildeverythinga full calibration set with reference values on the new instrumentdifferent technology, different presentation, or transfer fails validation
Fixing predictions with bias and slopeInteractive · simulated data
Points are transfer samples predicted on the new instrument with the master calibration. A bias correction moves the cloud onto the line; a slope correction also fixes a tilt. Neither can fix scatter, which is the sign that spectral standardization or a rebuild is needed. Simulated; numbers are illustrative.

Which samples should you use for the transfer?

Transfer samples should look like your real product and cover its full range: low and high values, every main ingredient source, every season you expect. Pick them for spread, not convenience. Algorithms such as Kennard-Stone choose samples that cover the spectral space evenly, which is better than taking whatever was on the bench that day. The samples must be stable, because they are scanned on both instruments.

A tray of small glass jars holding different grain and feed samples in neat rows
A transfer set: stable samples chosen to cover the full range of products and values, scanned on both instruments. Illustration, not a specific product.
Watch out

A transfer set from the middle of the range hides slope errors. If every sample sits near the average, a tilted correction looks perfect on the transfer set and fails on your highest and lowest loads.

How do you prove the transfer worked?

Use samples that were not in the transfer set, with fresh reference values, measured on the new instrument. Compare the new instrument's bias and prediction error with what the master achieves on similar samples. If they are close and the residuals show no pattern across the range, the transfer holds. If not, step back to diagnosis rather than adding another correction on top.

Then keep watching. Send a few routine samples to the reference lab every week from each instrument, plot the differences, and act when the bias trend moves. Lamp changes, repairs and relocations are all reasons to re-check.

Should you transfer the calibration or rebuild it?

Transfer or rebuild?Interactive · decision guide
A rule-of-thumb guide based on the methods above, not a guarantee. Whatever it says, the validation step decides.

If you are updating a calibration on an existing analyzer rather than moving it, the same logic applies in time instead of between boxes: add new, well-referenced samples that represent the change (a new supplier, a new crop), rebuild or extend the model, and validate on samples it has not seen. A custom feed calibration built from your own ingredients usually outperforms a generic one for exactly this reason.

Try this first

Scan five of yesterday's samples on both instruments today. If the predictions agree within your tolerance after matching presentation, you may need only a bias check. If they scatter, plan a standardization or a rebuild.

Try it free

After the transfer, run your transfer samples on the second instrument and enter the reference standard deviation and the new SEP in the free RPD and SEP calculator. If RPD drops below about 3, the transfer is not good enough for routine QC yet.

Worked example: a second site, same model

ILLUSTRATIVE A made-up company to show the reasoning; not a customer case.

A feed company opens a second mill and buys the same NIR model it already runs at the first site. The master calibration for protein in finished feed reads consistently high at the new site. The team first checks the obvious: the new site uses a different sample cup. After switching to the same cup type, most of the offset disappears, but a small tilt remains: low-protein feeds agree, high-protein feeds read high.

They pick transfer samples across the full protein range, scan them on both instruments, and compare spectra. The difference is smooth with a slight wavelength shift, so they apply piecewise direct standardization rather than a simple bias. A separate validation set, with fresh reference values, now shows the new site matching the master. They schedule weekly reference checks at both sites and a re-check after any service visit.

Quick check

1. After a bias correction, low-protein samples read low and high-protein samples read high. What is left to fix?
2. The second instrument uses a different sample cup. What should you do first?
3. How should transfer samples be chosen?
Answer the three questions above to see your score.
Where you are on the Instrument path

Step 4 of 4: Move a calibration

4 of 4 steps
  1. 1Inside the instrumentLight source, sample interface, detector.Read part 1 →
  2. 2Choose the technologyFT-NIR, dispersive or filter.Read part 2 →
  3. 3Benchtop, at-line or process?Where the instrument should live.Read part 3 →

Frequently asked questions

What is calibration transfer in NIR spectroscopy?

Making a calibration built on one instrument give the same results on another. It usually means scanning a set of samples on both instruments and correcting the difference, then validating on new samples.

Why do two identical NIR instruments give different results?

Small differences in wavelength axis, bandwidth, detector response and sample presentation change the spectrum slightly, and the calibration's wavelength weights turn those differences into different numbers.

What is the difference between slope/bias correction and PDS?

Slope/bias correction adjusts the predicted values. Piecewise direct standardization (PDS) adjusts the spectrum itself, wavelength by wavelength, so it can handle small axis shifts and bandwidth differences that a bias cannot.

How many samples do you need for calibration transfer?

Enough to cover the full range of values and products the calibration must handle. Spread matters more than count; choose samples for coverage, for example with the Kennard-Stone algorithm, and validate on separate samples.

How do you update an NIR calibration on an existing analyzer?

Add new samples with good reference values that represent the change, such as a new supplier or crop, extend or rebuild the model, and validate it on samples it has not seen before using it routinely.

Can you transfer a calibration between FT-NIR and dispersive instruments?

It is possible with spectral standardization, but differences between technologies are larger, so it often fails validation. Plan for a rebuild and treat a successful transfer as a bonus.

Next questions people ask

NIR Fundamentals course · 32 lessons

Keep every instrument telling the same story

The course covers calibration building, validation, transfer and routine checks across instruments and sites, in 32 short lessons for food, grain, feed and dairy labs. One payment, lifetime access.

Sources

  1. Workman, J. J. "A review of calibration transfer practices and instrument differences in spectroscopy." Applied Spectroscopy 72, no. 3 (2018): 340–365. doi:10.1177/0003702817736064
  2. Feudale, R. N., Woody, N. A., Tan, H., Myles, A. J., Brown, S. D., Ferré, J. "Transfer of multivariate calibration models: a review." Chemometrics and Intelligent Laboratory Systems 64, no. 2 (2002): 181–192. doi:10.1016/S0169-7439(02)00085-0
  3. Burns, D. A., Ciurczak, E. W. (eds.). Handbook of Near-Infrared Analysis, 3rd ed. CRC Press, 2007.
  4. Pasquini, C. "Near infrared spectroscopy: a mature analytical technique with new perspectives — a review." Analytica Chimica Acta 1026 (2018): 8–36. doi:10.1016/j.aca.2018.04.004
  5. Rinnan, Å., van den Berg, F., Engelsen, S. B. "Review of the most common pre-processing techniques for near-infrared spectra." TrAC Trends in Analytical Chemistry 28, no. 10 (2009): 1201–1222. doi:10.1016/j.trac.2009.07.007