The single best predictor of whether a process analytics project succeeds is whether it runs a feasibility study before the capital request. Studies that skip feasibility produce two outcomes with roughly equal frequency: an analyzer that works and an analyzer that does not. Studies that include a structured feasibility step produce the first outcome with closer to nine-out-of-ten reliability.

This guide describes a feasibility study scoped to answer the question can this measurement be made on this process with this technique — and, ideally, how well. It is six steps; a competent team can complete it in three to six weeks; the cost is low single-digit percent of an analyzer’s installed cost.

Step 1: Define the measurement

Write the measurement specification in one paragraph. What is being measured, on what stream, with what cycle time, with what required precision, against what reference method, with what consequence if the measurement is wrong by X.

This sounds trivial. It is not. Most feasibility scoping problems trace back to a measurement specification that uses unspecific words: concentration (of what, in what units, against what reference?), real-time (sub-second, sub-minute, sub-hour?), good enough (RMSEP target in absolute or relative terms?). A measurement specification that does not survive a five-minute interrogation is a measurement specification that will not survive a five-month project.

The artifact is a one-page document. The named author owns it; deviations require their sign-off.

Step 2: Map the variability

A chemometric model needs to be trained on data that span the variability the model will encounter in production. Step 2 is to enumerate the variability sources and assess whether the planned feasibility samples cover them.

The list, by category:

  • Recipe variation: deliberate variation across product variants, scale-up campaigns, and recipe revisions.
  • Raw material variation: lot-to-lot variation in feedstocks, especially for natural-origin or supplier-dependent inputs.
  • Process variation: the within-batch and batch-to-batch swing in process parameters that produces in-spec but not identical batches.
  • Equipment variation: differences across equipment trains, probe revisions, instrument calibration cycles.
  • Environmental variation: temperature, humidity, line vibration, EMI from neighboring equipment.

A feasibility study that samples only one batch on one line covers maybe a quarter of the variability the production model will face. A feasibility study that samples five batches across two lines and three raw material lots starts to be representative.

Step 3: Plan the sample matrix and the reference method

The sample matrix is a small experimental design. Common shapes:

  • Spike-and-recovery: known additions of the analyte to a base matrix, covering the concentration range of interest. Useful when the analyte and matrix are both available cleanly.
  • Process-derived gradient: samples taken at multiple time points across a real process run, where the analyte concentration changes naturally. Most representative; usually the right design for a process measurement.
  • Library: a structured set of representative samples from the process archive, retrospectively analyzed. Useful when historical samples are stored properly.

For each sample, a reference value is needed. The reference method must be:

  • more precise than the target measurement; rule of thumb, three times more precise;
  • traceable to a recognized standard;
  • available at the cadence required (the rate-limiting step in most feasibility studies is reference-method capacity).

If the reference method is a wet-chemistry assay that takes four hours per sample, plan accordingly. The most common feasibility-study delay is reference-method backlog, not spectroscopy.

Step 4: Run the spectroscopy under conditions that resemble production

A feasibility study run in a benchtop laboratory under perfect conditions will systematically overstate what is achievable on the line. The study should run, where possible, under conditions that include the failure modes the line will impose:

  • Probe orientation and access matching what the line will allow;
  • Sample temperature matching process conditions;
  • Sample matrix including the contaminants and intermediates present in the process, not just the pure analyte and pure solvent;
  • Acquisition timing matching the cycle time the model will need to respect.

Some of these constraints are best deferred to a pilot phase; some can be approximated in feasibility. The goal is to identify failure modes early enough to inform the design.

Step 5: Build a preliminary chemometric model

A feasibility model is not a production model. The deliverable is a go/no-go prediction performance number — RMSEP, R², classification accuracy — under preprocessing and modeling choices that are simple, defensible, and likely to generalize. PLS with standard preprocessing (SNV, derivatives, mean centering) is the typical starting point.

Two outputs from this step matter:

  1. The performance number, with its confidence interval given the small sample size. A single-sample test set is not a confidence interval; bootstrap or cross-validation confidence bounds are.
  2. The principal-component analysis of the spectra and residuals, identifying systematic structure that the simple model has not captured. This frequently surfaces matrix variation, contamination signatures, or unexpected chemistries that change the project scope.

The model is not validated in this step. Validation is a separate exercise on a separate dataset after design freeze.

Step 6: Write the go/no-go report

The feasibility report has six sections:

  1. Measurement specification (from Step 1, possibly revised).
  2. Variability map (Step 2) and an explicit gap analysis: what was sampled vs. what production will see.
  3. Reference method with precision, traceability, and capacity.
  4. Spectroscopy results: example spectra, identified peaks or features, principal-component analysis.
  5. Model performance: RMSEP, prediction-vs-reference plot, confidence bounds, residual analysis.
  6. Recommendation: go, no-go, or scope change. If go, with the explicit assumptions about variability coverage that the production project must verify.

A go recommendation that does not enumerate the assumptions about variability coverage is a recommendation that has not been thought through. A no-go recommendation should describe the alternative paths the project should consider — different technique, different measurement point, different chemistry workaround.

What feasibility does not establish

Feasibility tells you whether the measurement is possible. It does not tell you whether:

  • The model will be stable for a year of production.
  • The probe will survive the actual cleaning regime.
  • The site team will keep the analyzer maintained.
  • The cost-benefit will be positive at scale.

These are pilot-phase questions. A good feasibility study tees them up; it does not pretend to answer them.

Cost calibration

A typical feasibility study, scoped as above:

  • 3–6 weeks elapsed time.
  • 2–10 hours of process-engineering time per week.
  • 5–30 samples with reference analyses (the reference-analysis cost dominates the budget).
  • One spectroscopy laboratory engagement (internal or contracted), 1–3 days of acquisition.
  • One chemometric analyst, 1–2 weeks of model-build work.

The total cash cost of an externally contracted feasibility study runs €15,000 to €60,000 depending on sample count and matrix complexity. The cost of a project that skips feasibility and discovers in pilot that the technique was a poor fit runs five to twenty times that.

The math is unambiguous.