Chemometric Calibration Drift & Recalibration

Process Controls

Baseline drift rate 2.0

Sensor noise 1.0

Line speed (batches/s) 2.0

No recalibration yet — model is running on its initial calibration.

Live Readouts

RMSEP

0.000

Current bias

0.000

Batches since recalibration

How it works

In Process Analytical Technology, a chemometric model turns a raw spectroscopic signal into a real-time estimate of a quality attribute — here, a reactant concentration read from the peak absorbance at one analytical wavelength. A simple univariate calibration is:

A(t) = s·C_true(t) + b(t) + ε(t)
Ĉ(t) = (A(t) − b₀) / s

s   = calibration slope (assumed known, fixed at 1)
b₀  = calibration intercept, fixed at the last recalibration
b(t)= true instrument/baseline offset, which drifts over time
ε(t)= sensor noise

As long as the real baseline b(t) matches the calibrated intercept b₀, the prediction Ĉ is unbiased. But raw-material variability, fouling, and optical drift move b(t) away from b₀ over time, so the model develops a systematic bias = (b(t) − b₀)/s, even though the sensor itself is working fine — this is exactly the "Models must be maintained and updated" problem regulators flag under ICH Q8–Q10.

Two numbers track model health:

Bias(t)   = Ĉ(t) − C_true(t)
RMSEP     = sqrt( mean[ (Ĉ(t) − C_true(t))² ] )  over the visible window

Baseline drift rate — how fast b(t) wanders away from the last calibration point (raw-material aging, sensor fouling).
Sensor noise — random measurement noise ε(t) added to every scan.
Line speed — how many simulated batches run per second.
Pull Reference Sample & Recalibrate — takes one noisy lab reference measurement of C_true and re-solves b₀ so the bias collapses back toward zero, exactly like a periodic wet-chemistry check keeping an in-line NIR/Raman model honest.

The waterfall grid is the simulated spectrum scrolling through time (wavelength channel on one axis, absorbance as bar height); the two glowing traces above it are the true concentration (green) and the model's live prediction (amber), with small markers where a recalibration was performed.

True concentration Model prediction Recalibration event

⚠ Couldn't load the 3D engineThree.js failed to load from the CDN. Check your connection and reload.