We use cookies to ensure the site functions properly, measure traffic, and personalize marketing only with your consent.
Spectrally OS · Predictive models · Beta v1
Our stack deliberately combines CNN (convolutional neural networks) and classical chemometricsPLS, PCA) — selected for the specific measurement case. This material demonstrates the mechanics.
PLS and PCA assume an approximately linear relationship between the spectrum and the measured process variable.
Below are two independent sources of evidence: (A) aggregated metrics from a broader test setpls_vs_cnn_metrics.csv) and (B) one specific production run with 7 client reference measurementslepkosc_2026-06-01.csv), showing exactly where the divergence begins.
Metrics computed on the full test set, independent of the run in Section B. Each row shows the same.
The same viscosity, the same process moment, three independent sources: two models computed in parallel from the.
For each of the 7 moments at which the client performed an independent reference measurement, we compare the nearest.
| Time | Reference | CNN | CNN error | PLS | PLS error | Closer to reference |
|---|
Across the 7 checked points, PLS is closer to the reference once — at the lowest viscosity (approx. 90 units, process start,.
The difference is not constant — it grows together with the nonlinearity of the process. This is an argument for a hybrid stack, not for.
Linear range (process start)
At low viscosity (approx. 90 units), PLS and CNN produce comparable, small errors — PLS was even slightly.
Nonlinear range (middle and end of run)
Above approx. 350 units, PLS error grows faster than CNN error — on the full test set, the difference reaches.
Conclusion for the stack
We match the model component to the nature of the process, not the other way around. This run is evidence of the selection.