A rapid optical prediction and a chemical water determination can report the same units while relying on different evidence. Near-infrared moisture models are often built against reference measurements. The key comparison is therefore not simply speed: it is whether the model remains valid for the formulation, physical form and water range of the new sample.
Identify which result supplies the calibration target
In a 2010 freeze-dried-sample study, near-infrared spectra were related to Karl Fischer water results using multivariate analysis. The model used spectral information to predict the reference quantity rather than independently counting water molecules.Towards robust water-content determination by near-infrared spectroscopy (opens in a new tab)
This creates an evidence relationship: reference results support the calibration, and the calibration supports later predictions within its evaluated scope. An error in the reference values can therefore matter even when the optical measurement is repeatable.
| Check | Question |
|---|---|
| Calibration fit | Does the model describe the examples used to build it? |
| Prediction test | Does it perform on examples withheld from that fit? |
| Transfer assessment | Does it still perform after a relevant material or setup change? |
Treating these checks as interchangeable can make a model appear more widely established than the evidence permits.
Read a transfer result with the tested materials attached
The 2010 study built a model from mannitol–sucrose samples and predicted water in additional samples containing proteins, excipients or lower-density solids. It reported a prediction error of 0.19% in that evaluated set and explicitly limited its conclusion to the investigated conditions.Towards robust water-content determination by near-infrared spectroscopy (opens in a new tab)
The useful finding is that transfer was tested, not merely assumed. It does not follow that any different research-peptide powder belongs to the same validated prediction domain.
For a new material, ask which differences were represented in the study and which were absent. A changed component, physical appearance or concentration range may require evidence that the model still recognises water rather than a correlated feature.
A model can work across some compositional variation without working across all variation. The boundary is an experimental finding to document, not a contradiction.
A second study makes physical-form limits explicit
Clavaud and colleagues developed moisture models for monoclonal-antibody lyophilisates using designed variation in formulation and freeze-drying parameters. The published model excluded completely collapsed samples and moisture contents above 6%.Clavaud and colleagues — Moisture-model generation by designed experiments (opens in a new tab)
The authors described transfer within tested moisture and protein-to-sugar ranges, with a macroscopically intact cake. Those restrictions belong with the proposed use of the model.Clavaud and colleagues — Moisture-model generation by designed experiments (opens in a new tab)
They are not general acceptance limits for Novum products. In particular, an intact-looking cake does not establish an acceptable water content, and 6% is not supplied here as a universal moisture specification.
Compare the evidence package, not just the turnaround time
A useful method comparison includes the reference method, calibration materials, prediction samples, error measure and scope of any transfer exercise. It should also identify which samples would be referred back for another measurement.
For an illustrative decision, a new formulation matching the tested range and appearance may justify investigating model applicability. A visibly different physical form outside the studied range cannot be accepted solely because the software produces a number.
The optical route can add useful measurement capacity when its model is supported. Its convenience does not remove the need to understand what the number represents or how far the supporting comparison extends.
Sources and further detail
- Towards robust water-content determination by near-infrared spectroscopy (opens in a new tab)
Complete indexed original 2010 publisher abstract read after direct retrieval failed. Calibration/remaining-sample distinction and 0.19% prediction error retained with the tested scope.
- Clavaud and colleagues — Moisture-model generation by designed experiments (opens in a new tab)
Complete original abstract read; published online 2019, January 2020 issue. Excluded collapse/high-moisture cases and transfer boundaries retained, not converted into product specifications.
Sources checked 20 September 2026. Worked examples are illustrative unless a supplied report is explicitly identified. This article has not undergone independent scientific peer review.