Two cases from how the model is built that deserve more room than a paragraph: a rheological property that no single band encodes, and a real production mixture whose components overlap in the same spectrum.
Scroll
Purely classical models are cheap to validate and blind to non-linearity. Purely neural models absorb non-linearity and demand data. Real process chemistry needs both, chosen case by case.
Full-spectrum input, non-linear response, overlapping bands, multi-component mixtures, and derived properties such as viscosity that no single band encodes.
Compact calibrations from a limited number of samples, loadings a chemist can inspect and argue with, and quick validation against titration or HPLC.
Faster, more accurate models that generalise better than a single-family approach. In a feasibility study on alkyd resins, a CNN over the whole spectrum outperformed linear regression on the same data, reaching R² above 0.99 for both acid value and viscosity.
The most common objection we hear is that Raman measures composition, not properties. It is a fair objection, and the answer is a chain of physics you can follow link by link — not a black box.
At the start of a polycondensation the vessel holds mostly free monomers. Nothing about that mixture is viscous yet, and no single chemical concentration would tell you what the batch will pour like at the end.
As the reaction proceeds, monomers link into chains and the distribution of chain lengths shifts. Viscosity follows that distribution — and so do the molecular vibrations, because the bonds themselves are what changed.
Many adhesives and resins are made by polycondensation, and some polymerise while curing. That is exactly what we see — the change in chemical bonds. dr Bartosz Kawa, CTO · Główny Mechanik
Overlay the spectrum from the start of the batch on one taken later and the peak positions barely move — the ratios between them do. That shift is small, systematic, and far too subtle to read off a chart by eye.
A CNN reads every channel rather than a few chosen peak windows, with classical chemometrics alongside it where a compact, inspectable calibration wins. The model is fitted on this plant’s own samples against its own reference values.
In a documented feasibility study on alkyd resins the model reached R² above 0.99 for viscosity through the whole synthesis, at a prediction uncertainty of ±0.25 Pa·s — and the acid value came out of the same spectrum at about ±0.2 mg KOH/g, without a titration.
A production sample is a mixture, and the bands of its components sit on top of one another. This is the case where a single-parameter sensor gives up and a model does not.
A pure substance has a clean fingerprint: a handful of bands in known positions with known relative intensities. Measured alone, in a laboratory, it is a solved problem.
Add three more and the bands start sharing space. Some overlap almost completely, and the intensity at any given wavenumber now belongs to more than one substance at once.
The instrument never sees the components separately. It sees their sum, plus whatever the matrix adds. Reading a peak height off this curve would give you a number that belongs to no single component.
Calibration uses real process mixtures with reference values, not a library of pure substances. The model learns how the overlap behaves as the composition moves — including the non-linear part, which is where classical single-band approaches break down.
This is the everyday case in our feasibility work: silicone contamination in PA66 recyclate quantified across a 0.3–1.0 % calibration range, and ammonium nitrogen predicted in industrial water even though its signal overlaps the water bands.
Feasibility runs on real process material — a rheological property, a multi-component blend, or both. The report says what came out, at what uncertainty.