Machine-learning models that infer a hard-to-measure variable — composition, quality, emissions — from cheaper, faster process signals already being collected.
Embedded-ML spectroscopic & chemometric analysers
Field Instrumentation
Industry Acceptance
Near-term · 2–5 yrs
Evidence E3 · First production use
AI / ML
In-line NIR/Raman/FTIR analysers with the chemometric model embedded on-device, giving real-time composition data without a separate host-PC modelling step.
Short entry. This page summarises our current placement. Full, sourced analyses are published first for the technologies closest to practical use; see the list of full analyses.
How it works
A spectroscopic analyser shines light through or onto a process stream and records how much is absorbed or scattered at each wavelength. Near-infrared, mid-infrared (FTIR) and Raman instruments each see different molecular features. The raw spectrum is not a composition; it has to be converted by a calibration model.
Chemometrics is the multivariate mathematics that does the conversion. Spectra are pre-processed to remove baseline and scatter effects and then passed through a regression model, classically partial least squares, built from samples with known laboratory values. Embedding that model in the analyser itself removes the separate modelling computer and delivers a composition value directly to the control system.
Why it matters
Shrinks the gap between lab-grade composition analysis and real-time control, particularly valuable in specialty chemicals, pharma and food & beverage.
Where it fits
The approach is established in pharmaceutical manufacturing under the process analytical technology framework, and in polymers, specialty chemicals and food, where a composition or end-point is needed in seconds rather than hours. It is most valuable where the measurement can close a control loop or release a batch earlier.
Why we place it at the Industry Acceptance stage
Established in pharma (PAT) and specialty chemicals; extending into bulk chemicals and refining as embedded compute cost falls. Estimated time to the Proven in Use stage: 2–5 yrs. What the stages mean.
Key risk
Model recalibration for feedstock or product changes still requires chemometric expertise most sites don't have in-house.
Questions to ask before committing
- How many reference samples, spanning what range of conditions, were used to build the calibration?
- What is the procedure when feedstock or product grade changes, and who is qualified to carry it out?
- Can a model be transferred to a replacement or second instrument without starting again?
- How does the analyser signal that a spectrum looks unlike anything in its calibration set?
Standards and terms to know
- PAT
- Process analytical technology, the regulatory framework that encourages in-process measurement of quality attributes.
- PLS
- Partial least squares regression, the workhorse calibration method in chemometrics.
- Calibration transfer
- Making a model built on one instrument valid on another.
Related technologies
See this entry on the full maturity map · All Instrumentation entries · How placements are made. Stage and time-horizon placements are editorial judgements based on public information as of October 2026, not engineering, safety or investment advice; see the Terms of Use.