From the isolated NCO band to a value you can act on.
How an immersion Raman probe in the reactor - measuring inline, directly in the medium with no sample drawn - turns the isolated isocyanate NCO band, the urea breach band and the epoxy endpoint into NCO%, the R index, EEW and residual on your control system in seconds - and keeps every operator away from the diisocyanate sample.
The reactive endpoint maps onto a handful of specific bands.
Shine monochromatic laser light into the reaction mixture. Most of it scatters back unchanged; a small fraction exchanges energy with molecular vibrations and returns shifted, and those shifts are characteristic of the bonds that caused them. The isocyanate NCO stretch at ~2270 cm⁻¹ is one of the strongest, most isolated targets in vibrational spectroscopy, so NCO%, the epoxy endpoint and residual monomer come back cleanly from one spectrum.
Inelastic scattering carries the information
The laser (785 or 1064 nm) illuminates the medium through the immersion probe. Photons that scatter inelastically shift in wavelength by amounts set by the molecular vibrations they met. Plotting intensity against Raman shift (cm⁻¹) gives the raw spectrum.
The bands these chemistries run on
NCO% reads from the strong, isolated band at ~2270 cm⁻¹, and the R index (NCO/OH ratio) follows it. A urea band appearing at ~1630 cm⁻¹ flags a breach of the reactor's dry atmosphere. The epoxide equivalent weight and residual epichlorohydrin read from the same spectrum. Each band is narrow and characteristic, so several parameters resolve from a single read.
Composition, not a proxy
Raman reads composition directly. An inline viscometer infers a reaction-progress proxy but no NCO%; NIR reads NCO too, but on a less isolated band than the Raman ~2270 cm⁻¹. The Raman spectrum carries NCO%, the endpoint and residual monomer at once.
NCO titration and GC remain the analytical reference. Inline Raman replaces them where decision time, measurement continuity and operator safety matter - the offline methods become verification, not the bottleneck.
From the reactor wall to the control room.
Four links in the chain: an immersion probe in the process, fiber optics to the analyzer, chemometric models in Spectrally OS, and your control system at the end. There is no sampling loop anywhere in it - the measurement is inline, taken directly in the medium.
Immersion probe in the reactor or line. IP67, ATEX/IECEx Zone 0, fitted with the self-cleaning module for reactive resins that film or cure on optics.
Excitation light travels to the probe and scattered light returns over fiber, so the analyzer can sit away from the process area while the probe stays in the medium.
785 / 1064 nm laser excitation, back-thinned TEC-cooled CCD detector. Up to 2 measurement channels as standard, more on request.
AI and advanced chemometrics with a CNN core translate the spectrum into NCO%, the R index, epoxide equivalent weight and residual values plus PASS-FAIL flags, then hand them to your PLC / DCS.
The hardware in the chain
The numbers that define the measurement path, probe to analyzer. Calibration is automatic, against a reference integrated in the probe - there is no manual recalibration routine to schedule.
| Measurement principle | Raman spectroscopy |
| Laser wavelength | 785 nm / 1064 nm |
| Detector | Back-thinned CCD, TEC-cooled |
| Acquisition time | 5-300 s |
| Fiber optic reach | Up to 100 m |
| Measurement channels | Up to 2 (expandable on request) |
| Probe | Spectrally X1 PROBE, IP67, ATEX/IECEx Zone 0, self-cleaning |
| Calibration | Automatic - reference integrated in probe |
| Communication | PROFIBUS, PROFINET, GSM |
Raman is the core, not the whole toolbox
The core technique and the workhorse here - the isolated NCO band at ~2270 cm⁻¹ is a near-ideal Raman target, with a CNN doing the interpretation.
Surface-enhanced Raman on metallic nanostructures, typically Ag or Au - used when analyte concentrations sit at ppb-ppm levels, below classical Raman sensitivity.
Near-infrared as a complement where Raman bands are weak. It reads NCO too, but on a less isolated band than the Raman ~2270 cm⁻¹.
Complementary sensing for specific challenges. The technique is chosen for the chemistry, not the other way around.
Reactive resins cure on optical windows. The probe is built for that.
Epoxy, polyurethane-prepolymer and reactive-adhesive deposits film and cure on any optical surface left in the process. In strongly fouling media the Raman signal would degrade within hours to days of operation. The self-cleaning module removes that failure mode with a mechanical cycle: retract, rinse, return.
Why optics film
Curing epoxy and prepolymer layers build on the probe optics as the reaction runs. A filmed window means drifting spectra first, useless spectra later - in strongly fouling media within hours to days.
The cycle preserves measurement geometry
The probe retracts from the medium, the optical window is rinsed, and the probe returns to the same measurement position. Geometry stays fixed, so readings stay comparable across the whole campaign.
The probe stays in, the operator stays out
Because the retractable probe cleans itself in place, no one opens the reactor to draw a manual TDI/MDI sample. The probe runs ATEX/IECEx Zone 0, on a 30 mW laser where the zone requires it.
A capability, not an accessory
Without self-cleaning, a process probe in reactive-resin service needs manual cleaning - and manual cleaning means breaking a continuous measurement. The module is integrated with Spectrally X1 PROBE and Spectrally X1 INLINE, and it is what makes continuous 24/7 measurement realistic in epoxy, prepolymer and reactive-adhesive chemistry.
It also closes the exposure route. The cycle runs automatically during production, so there is no manual optics service in the middle of a campaign, no gap in the data your control system sees - and no operator at the reactor drawing a diisocyanate sample.
Spectrum to a process value, continuously.
Up to 40 minutes of di-n-butylamine NCO titration wait, gone.
Trace components resolved down to single ppm across the platform.
From feasibility to a live inline measurement, 3 to 5.5 months.
Spectrally OS turns spectra into operating information.
Raw spectra are useless to an operator. Spectrally OS translates them into concentrations, ratios, PASS-FAIL classifications and alarms - and keeps a logged, auditable history of the models doing the translating.
CNN core - the workhorse
A convolutional network reads the whole band structure at once - is this prepolymer on its NCO target, does this incoming polyol match the specification. On reactive-resin spectra the CNN core outperforms the PLS/PCA benchmark it is measured against.
Regression to numbers
The model maps the spectrum to numbers - NCO% and the R index, epoxide equivalent weight, residual epichlorohydrin and free isocyanate monomer - several parameters from a single read.
Anomaly detection
The model reduces a spectrum to its drivers and flags what does not belong - a urea band appearing at ~1630 cm⁻¹ surfaces as a dry-atmosphere breach, not as a customer return.
Models start from feasibility - typically your process data plus samples measured in the Gekko lab. Built on your epoxy or isocyanate matrix, not a generic library.
Predictions are checked against your reference method - di-n-butylamine titration for NCO, GC for residual monomer - before the model is trusted in production.
Models are reinforced with your production data as it accumulates. Updates deploy without stopping the analyzer or the line.
Scheduled when the recipe, raw material or process parameters change. A model change - not a platform redesign.
Spectrally OS at a glance
One model layer runs across Spectrally X1 LAB, X1 PORTABLE and X1 INLINE, so a model built during feasibility migrates to the process instrument. Routine operation uses no reagents and no consumables.
| Supported instruments | Spectrally X1 LAB / PORTABLE / INLINE |
| Predictive models | AI / advanced chemometrics, CNN core (PLS/PCA benchmark) |
| Spectra database | Proprietary + public, ~28,000 spectra |
| Operator interface | SpectrallyUI |
| Operating system | Debian GNU/Linux 13.2 |
| Calibration | Automatic, from a built-in reference signal |
| Data export | CSV, PDF, RAW |
| User management | Role-based access control (RBAC) |
| Data storage | Local database - logging continues without a network |
| Model updates | Deployed without stopping the analyzer |
A value counts when it lands in your control system.
Spectrally streams process values - not raw spectra - into the plant's control and data layers. SpectrallyUI shows trends and alerts at the analyzer; the same values reach your PLC / DCS, where a reading can drive a setpoint correction, an alarm, a quality report or a batch hold.
PLC / DCS
Values arrive as process signals over PROFIBUS or PROFINET - unified across all Spectrally products. To the operator, the analyzer reads like any other instrument on the loop.
MES / SCADA / LIMS
Configured per site: batch context in MES, trending and supervision in SCADA, reconciliation with lab records in LIMS.
GSM and export
GSM is available across the product line alongside the fieldbus protocols. Data exports in CSV, PDF and RAW for reporting and archive.
Network fit
Segmentation, white-listing, no open ports to the corporate network - typical OT policies are scoped with your automation and IT teams during integration.
Industrial cyber
IEC 62443 expectations for industrial automation systems are addressed in integration scoping, together with backup and recovery of measurement data.
Time-stamped batch record
Measurement and logging run on a local database, so a continuous, time-stamped composition record accrues per batch - and with the manual sample eliminated, that record stands in for an operator exposure. Model changes are logged and auditable.
Run the full chain on your own reactor.
A feasibility study starts with representative samples or process data from your epoxy or prepolymer stream. Gekko builds the chemometric model on your matrix, benchmarks it against your reference method - di-n-butylamine NCO titration or GC - and hands back the spectra, band assignments and the validated numbers behind your NCO%, R index and EEW questions, before any hardware decision.