Spectra Lab: chemometrics software for near-infrared calibration, in your browser

Interactive demo: spectral preprocessing, PLS regression and model validation

Choose a preprocessing method and watch what it does to the spectra, the RMSECV-versus-components curve, the predicted-versus-reference scatter, and the validation metrics. Start with raw spectra, then try SNV: on this real meat data it lowers the error, then compare the derivatives.

Results computed by the Spectra Lab engine on a real public meat dataset (172 training samples, fat % as-is), 5-fold cross-validation.

Tour the Lab

Six screens take you from a raw CSV to a filed validation report. Pick a step on the left; on a phone the tabs scroll sideways.

1. Import 2. Preprocess 3. Calibrate 4. Validate 5. Report 6. Second opinion

Upload a CSV with one row per sample and your reference values alongside the wavelengths. Spectra Lab previews the table so you can check the columns, the row count and the reference range before anything is computed.

Learn more: importing spectra into Spectra Lab

Compare raw and preprocessed spectra side by side before you train anything. Standard normal variate, multiplicative scatter correction, Savitzky-Golay derivatives and baseline correction are all applied the same way to every spectrum, so you can see the effect instead of guessing it.

Learn more: preprocessing and scatter correction

Build a PLS regression model against your reference values with k-fold cross-validation. The number of components is chosen automatically as the count with the lowest RMSECV, so the model is not tuned by hand to flatter itself.

Learn more: PLS regression and component choice

One view holds RMSECV, R2 CV, RPD and bias together with the predicted-versus-reference plot and the residuals. Reading them together is what tells you whether the error is even across the range or hiding a systematic bias at one end.

Learn more: validating a calibration

The report screen shows the cross-validation-by-factor curve next to the numbers behind it. From here you export a PDF validation report you can file with the method documentation, so the model you signed off is the model you can show an auditor.

Learn more: what belongs in a validation report

Score new spectra against a model you already trained and get Hotelling T2 and Q residual flags for every sample. T2 tells you which samples sit far from the calibration centre; Q tells you which ones the model cannot reconstruct, which is usually the sign of a bad spectrum or a different product.

Learn more: outlier detection with T2 and Q

Real datasets, real results

We ran real public NIR datasets through Spectra Lab, the same engine students use, not just simulated teaching spectra. The results line up with published work, and each one teaches a lesson worth repeating yourself.

Data: mango dry matter from Anderson, Walsh, Subedi and Hayes (2020), Postharvest Biology and Technology 168, dataset CC BY 4.0 (Mendeley Data, doi 10.17632/46htwnp833), teaching subset of 600 + 100 spectra, 684-990 nm. Meat: public-domain dataset of 240 finely chopped meat samples, 850-1050 nm transmission, fat, moisture and protein by wet chemistry (reported as-is). Results computed by Spectra Lab, PLS with 5-fold cross-validation.

Learn by practicing

Spectra Lab is built for learning by doing. Every concept comes as a short exercise with a goal, the steps, what you should see and why it matters, and the workbench explains each step as you go.

1. Your first model. Import spectra, train PLS and read RMSECV, R2 and RPD.
2. When preprocessing hurts. Compare raw spectra with SNV and see a model fall apart.
3. Derivatives and baselines. Savitzky-Golay 1st vs 2nd derivative: baseline removed, noise added.
4. Get a second opinion. Score new spectra against a model and see which samples get flagged.
5. Try your own data. Import your CSV, train raw and one preprocessing, compare RPD.

Next, with the real datasets above: inside vs outside the calibration range, a new season, and as-is vs dry-matter basis.

Problems Spectra Lab solves

These are the questions analysts and statisticians keep asking about calibration. Here is what the Lab does about each one today.

Based on recurring questions from analysts and statisticians in public Q&A forums and the calibration literature.

Free chemometrics resources on SpectroScience

What you can do with Spectra Lab chemometrics software

From calibration curve to NIR calibration in near infrared spectroscopy

A single-wavelength calibration curve goes back to the Beer-Lambert law: absorbance is proportional to concentration, so one wavelength and a linear fit can quantify one analyte in a clean matrix. Real samples are not clean. Scattering, particle size, temperature and overlapping absorption bands break the single-wavelength assumption.

Multivariate calibration solves this by using hundreds of wavelengths at once and letting partial least squares regression find the combinations that carry the information and reject the rest. Spectra Lab walks that whole path for you: Beer-Lambert calculator, NIR calibration methods compared: MLR, PCR, PLS and ANN, and How to build a PLS model in practice: from raw spectra to validated calibration.

What is partial least squares regression? PLS vs principal component analysis (PCA), in plain words

Model validation like a lab manager: RMSECV, RPD, bias and Hotelling T2

A calibration is only as good as its validation. Spectra Lab reports the numbers a lab manager asks for, and lets you pull them into a report you can defend in an audit.

RMSECV

Root mean square error of cross-validation: the error you should expect on new samples from the same population, computed by leaving samples out and predicting them back.

RMSEP and bias

When you have an independent test set you also get RMSEP, and the bias tells you whether the model is systematically high or low rather than just noisy.

RPD thresholds

RPD is the ratio of reference standard deviation to prediction error. As a rule of thumb, below about 2.5 a model is only good enough for screening, and above about 3 it can be considered for quantitative use.

Hotelling T2 and Q outliers

T2 flags samples that sit far from the calibration centre; Q residuals flag samples the model cannot reconstruct. Together they separate unusual samples from bad spectra.

Read more: Cross-validation in NIR calibration: leave-one-out vs k-fold · Choosing the right NIR calibration validation approach · How to detect NIR spectral outliers.

Learning path: the course lessons that teach each step

Spectra Lab is the practice bench for the NIR spectroscopy course. Each stage of the workflow maps to a lesson that explains the theory behind it.

• Import and clean spectraLesson 25: Sample PreparationSee it in the curriculum →
• Preprocessing and scatter correctionLesson 21: Reading the NIR SpectrumSee it in the curriculum →
• PCA and score plotsLesson 22: What Is ChemometricsSee it in the curriculum →
• PLS regression and component choiceLesson 23: Introduction to CalibrationSee it in the curriculum →
• Validation, RMSECV, RPD and biasLesson 24: Validation TechniquesSee it in the curriculum →
• Outliers, T2 and Q diagnosticsLesson 31: Troubleshooting and Problem SolvingSee it in the curriculum →

New to NIR and chemometrics? Get learner access with the NIR course.

Read the articles behind each step

Field-tested explanations from the SpectroScience blog, each paired with something you can try in Spectra Lab.

• PLS Regression for NIR: Step-by-Step Guide for Food and Feed CalibrationRead the article →
• Why Does R² Lie? Understanding RPD, SEP, and RMSEP in NIR AccuracyRead the article →
• NIR Calibration Overfitting: Why It Happens and Three Validation MethodsRead the article →
• Why Do NIR Outliers Appear — and When Should You Remove Them?Read the article →
• Why Does Your Reference Method Limit NIR Accuracy? Lab Error Study ExplainedRead the article →
• As-Is vs. Dry Matter in NIR: Understanding the Key DifferencesRead the article →

Frequently asked questions

What is partial least squares regression?

Partial least squares (PLS) regression is a chemometric method that predicts a property, such as moisture or protein, from many correlated variables, such as the absorbances at every wavelength of a near-infrared spectrum. It finds a few latent components that explain the spectra and the property together, then regresses on them.

What is a Savitzky-Golay filter used for in spectroscopy?

It smooths spectra and calculates first or second derivatives by fitting a polynomial in a moving window. Derivatives remove baseline offsets and slopes, which often improves NIR calibration models.

Is Spectra Lab free?

The interactive demo on this page is free to use now. Student access to the full Lab is coming soon, and NIR course students will hear first.

Which instruments' spectra work with Spectra Lab?

It is instrument-independent. Any NIR analyzer or spectrometer whose spectra you can export as numbers will work, regardless of make or model.

What file format does Spectra Lab accept?

Spectra Lab accepts CSV files with one row per sample.

Is my data private?

Your uploads stay inside your own Spectra Lab account and are used only to run your calibrations; they are not shared with other users.

PLS or PLS-DA?

PLS regression predicts a continuous value such as moisture or protein, and that is what Spectra Lab does today. PLS-DA (discriminant analysis) predicts a class, such as pass or fail; it is on the Spectra Lab roadmap.

How many samples do I need?

As a working rule, start with at least 6 to 10 samples per PLS component you expect to need, spread across the whole range of the property you want to predict.

How many components should I use in a PLS regression?

Choose the minimum of the cross-validated error curve, and prefer fewer components when the curve is flat.

What if RMSEP or RMSECV has no clear minimum?

Pick the first component count after which the error stops improving meaningfully. A flat curve can mean the signal is weak or the preprocessing is unsuitable.

Should I mean-center or autoscale spectra before PLS?

Mean-centering is standard. Autoscaling gives noisy wavelengths equal weight and is usually avoided for spectra.

How much variance should PCA capture?

There is no fixed threshold. Use enough components to describe the systematic variation, checked with score and loading plots.

How do I know a new sample is within my calibration?

Check Hotelling T2 and Q residuals against the calibration limits; Spectra Lab flags both.

Can I cross-validate with replicate scans?

Keep all replicates of one sample in the same fold, otherwise the error looks better than it is.