AI for Paint Color Formulation in Manufacturing: RFP and FAT/SAT from Recipe to Release

A factory may have an approved color standard and still need several lab trials before its coating matches. Pigment lots change, another technician takes over, or the same formula produces a different appearance on a production part. For a Thai manufacturer considering AI for paint color matching and formulation, the purchasing question is specific: can the system turn spectrophotometer readings and approved raw materials into a usable candidate recipe, learn from a drawdown, recommend a controlled correction, and release only an approved formula to production?

This is a different problem from detecting color variation on finished parts. Our guide to AI inspection for color variation addresses the output inspection stage. AI prediction of paint film thickness addresses process settings and thickness. The present article concerns the physical formulation of paint: pigments, base, quantities, test panels, correction and authorization.

Conventional spectral formulation software has existed for years. Calling a product “AI” does not establish that it uses machine learning, let alone generative AI. X-Rite describes a workflow using measurement, available pigments and binders, trial samples, remeasurement, correction and approval. Datacolor describes formulation from spectral data and optional use of historical approved formulas. These vendor descriptions show available functions; their performance claims are not proof of results at your plant. A buyer should compare the proposed learned layer with a conventional formulation baseline on the same materials and test set.

Define the boundary between formulation and inspection

A formulation system returns a proposed composition and, after a test, a proposed addition or correction. Inspection decides whether a completed coating meets its specification. Both can use color coordinates and a spectrophotometer, but their operational outputs and accountable owners differ. Buying inspection software will not automatically solve recipe selection. Likewise, a film thickness model may help control appearance but will not determine which compatible pigments to mix. Put this boundary in the RFP before comparing vendors.

Assign people and rights to each step: who requests a color, who calculates candidates, who weighs the lab batch, who prepares the drawdown, who measures it, who appraises it visually, who approves the formula, and who may issue it to a dispenser or MES. A high model score is not an approval. Candidate, rejected, provisional and released recipes need separate states and a visible audit trail. The quality owner should be able to reject a numerically close formula when gloss, texture, opacity or durability is wrong.

Fix measurement and drawdown conditions before training a model

The input is not simply a Lab triplet. Record the target panel ID and condition, spectrophotometer geometry and aperture, calibration, specular component treatment, illuminant and observer settings. Record the substrate, primer, application method, wet and dry film targets, curing schedule, time until measurement, and gloss of the test panel. Identical liquid paint can look different when those conditions change. Do not compare a glossy production part with a matte laboratory panel using one color difference number and call the problem solved.

Plan repeat measurements. ASTM E1345-24 discusses reducing measurement variability with multiple readings or specimens when comparing a result with a tolerance. The suitable number of readings depends on panel uniformity and instrument repeatability; do not assume a universal count. If two measurements disagree, first check drying, panel identification, calibration and location of the readings. Only then investigate the recipe engine.

ASTM D2244-25 covers calculation of color differences and tolerances for instrumentally measured opaque specimens. Its public description stresses using the same color difference system, considering the direction of a difference and correlating numerical tolerances with visual acceptability. ASTM D1729-22 concerns controlled visual appraisal of opaque materials. Its public scope also notes that critical appraisal of metallic and pearlescent paints needs special directional illumination beyond its ordinary setup. Specify the actual lighting and angles that your customer uses; a reference to a standard number alone is not an acceptance method.

Define allowed recipes before asking for the “optimal” recipe

Supply the system with a controlled material master: permitted pigments, base types, concentrations, stock, cost, compatibility, minimum weighable addition, maximum loading, batch size and customer restrictions. A spectral target does not uniquely define a chemically acceptable formula. A visually close candidate can fail because its pigment is not approved, its base is incompatible, its concentration cannot be dispensed, or its coating properties have not been validated.

Separate hard constraints from ranking objectives. Hard constraints exclude an unapproved pigment, incompatible base, expired material, invalid mixing sequence or amount below the scale resolution. Ranking objectives may include predicted color difference, expected correction count, material cost and reuse of residual stock. A cheap candidate is still unacceptable if hiding power, weathering or process stability fails. Demand a reason for the candidate’s ranking and a clear display of violated constraints.

Ask vendors which component is physical spectral modelling, which is library search, and which learns from historical approved recipes or corrections. Datacolor’s Match Pigment literature describes a spectral formulation engine and SmartMatch using prior formulation knowledge. X-Rite describes candidate formulation and correction using available pigments and binders. Those are useful examples of possible workflows, not independent evidence that a specific AI model will work on your coatings. Require a side-by-side trial against a conventional formulation workflow with your colors, materials and drawdown method.

Link every candidate to its physical trial

A spreadsheet recipe name is too weak as a traceability key. At minimum, link target color ID, material system and version, pigment lots, candidate recipe ID, lab batch ID, drawdown panel ID, measurement conditions, result, correction step, approver and approval time. Preserve the relationship in both directions: which candidate produced this panel, and which panel supported this released formula? Keep rejected attempts and their reasons. A candidate may meet the numeric color tolerance but fail on gloss; a correction may be smaller than the dispenser resolution; a material may later be discontinued.

Version old data before it enters a learning system. Recipes made with obsolete pigments, old instrument settings or a different substrate cannot silently join the training set. Record effective dates, software and model versions, material master version and approval status. In a multilingual plant, keep pigment codes, color IDs and recipe states language independent; translate only display text. A free text note such as “make it less blue” is not an instruction to a dispenser. Send material code, exact quantity, target batch and approved recipe version.

Put gates between the lab batch and production

Stage integration. Initially, an experienced colorist reviews the suggestion and makes a small lab batch. Next, send an approved candidate to a weighing device or dispenser while recording the actual dispensed quantities. Later, issue a released recipe to production and connect its version with the batch and pigment lots. Test failure paths at each stage. “Dispenser integration” on a brochure might mean only exporting a file, whereas the plant needs prevention of duplicate commands, feedback of actual weights and recovery after a communication fault.

A lab success does not guarantee production success. Mixer shear, dispersion time, addition order, temperature, vessel size, batch size and application equipment can change the result. Keep both a proportional scale-up calculation and a measured pilot batch. After a correction, recalculate total batch mass, residual material and subsequent additions. Decide who may quarantine, rework, reuse or discard a batch when repeated additions threaten other properties.

Only a released version should reach production. Its evidence should include measured color under defined conditions, visual appraisal, drawdown preparation, required performance tests and any customer sign-off. Provisional approvals need expiry dates. A material master change should trigger a review before an old recipe is issued again. Never overwrite an approved historical version: otherwise the factory cannot explain which formula was used for a shipped lot.

Write an RFP around inputs, constraints and exceptions

Start with the actual coating families, number of matches per month, current measuring instruments, colorant database, scales, dispensers, application method, approval path and interfaces. If your current first-shot match rate is unknown, include baseline measurement in the project scope instead of inventing a percentage. Define “first-shot match”: does it require instrument tolerance alone, or instrument, visual appearance and performance together? Require the same definition in all vendor proposals.

Compare total project cost: instrumentation and calibration, colorant data preparation, historical data cleanup, lab trials, equipment interfaces, user training and maintenance. Vendor “match rates” may use different color sets, substrates, finishes and denominators. Treat catalogue figures as vendor claims; calculate your investment case from plant FAT/SAT evidence.

FAT/SAT must test how the system stops

For FAT, use representative target colors, a real material master and measured drawdowns. Include easy and difficult colors, light and dark shades, changed pigment lots, expired or prohibited pigments, noisy measurements and customer-specific tolerances. Document acceptance limits before tests. Do not impose one universal ΔE threshold on every coating family; customer specification and visual conditions determine tolerance. Log the magnitude and direction of color difference and the visual decision.

At SAT, repeat the workflow with plant instruments, operators, pigment lots, network and equipment. Disconnect the dispenser; retry a recipe transfer; revoke an approval; try a wrong panel ID, uncalibrated instrument, uncured sample, excess additive and stock shortage. The system should say “cannot judge” with a reason when data or constraints are inadequate. It should not replace missing data with a plausible number or send an unapproved recipe to the dispenser. Verify manual fallback and recovery without duplicate dispensing.

Use several metrics: first-shot match, number of corrections to release, elapsed approval time, candidate calculation time, discarded paint and panels, human work time and integration errors. Define each numerator and denominator by product family. A short pilot cannot reliably prove a reduction in long-term customer complaints; evaluate candidate quality and repeatability first. Preserve unsuccessful cases, because a dark color or a changed finish can fail while the overall average improves.

Compare three methods in the same PoC

Run the current expert workflow, a conventional spectral formulation baseline, and the proposed learned recommendation or correction layer with the same materials, measurements and test panels. Split evaluation by color family, pigment lot and target standard so the learned model cannot merely recall identical historical recipes. Compare desk predictions separately from actual weighed and sprayed batches. A recipe that forecasts a good color but asks for a quantity below scale resolution has failed the production test.

Track failures by stage. If repeated measurements disagree, improve measurement and panel preparation. If the same inputs return unstable recommendations, inspect model and master versions. If lab predictions are stable but pilot batches fail, investigate mixing, dispersion and application differences. Compare old and new model versions on the same retained test set before release. An improved overall average is insufficient when a key customer’s color family deteriorates.

For a Thai plant whose Japanese development team controls the standard and local production uses alternative materials, distribute the physical standard, spectral data, viewing conditions, tolerance and material approval under one version. Make the vendor’s system responsibility explicit when instrument, software and dispenser come from different suppliers. Contract for export of approved formulas, spectral readings and correction history in a usable format. Ask where recipes and customer colors are stored, whether cloud data can be reused for training, and how records can be removed at contract end.

Monitor the released system, not just the pilot

After go-live, review the share of suggestions actually used, the reason each rejected suggestion was declined, corrections per released color, time from request to approval, discarded paint and panels, integration faults and unauthorized-release attempts. Separate predicted color, measured lab color and measured production color in the dashboard. A high suggestion rate has little value if colorists always return to manual calculations; a high acceptance rate is concerning if visual appraisal has quietly disappeared. Break metrics down by coating family and customer standard instead of reporting one average across unlike colors.

Investigate a failed match in a fixed order. Check the physical standard, spectrophotometer calibration, pigment lot, actual dispensed weights, panel thickness and cure, measurement settings, recipe version and model version. Repeat a measurement when necessary. If the measurement system is unstable, fix it before retraining. If the lab is stable but the production tank fails, compare mixing shear, dispersion, addition order and application. Retain the old and new model versions and compare both on the same test set, including important difficult colors. Keep the owner of each corrective action visible.

Put a complete acceptance scenario into the contract

One practical FAT/SAT scenario begins with an approved target panel. The system measures it and proposes three candidates using only approved materials. Inject a prohibited but inexpensive pigment into the test material data and verify that no selectable candidate uses it. The colorist selects one candidate, weighs a lab batch, makes a drawdown, and imports its measured result. If outside tolerance, the system calculates an addition and predicted direction of change, but cannot send the instruction to a dispenser until a named user verifies the material and amount. After correction, a panel may meet the numerical tolerance but fail visual appraisal on gloss. The quality owner rejects it, and the reason remains in the history. Once approved, a material master change invalidates automatic issue of the old version and forces review. Repeat the numbered scenario during SAT with actual devices, operators and fault recovery, then compare its outcome with FAT.

Contract wording should specify what “connected to the dispenser” covers: export only, actual-weight feedback, wrong-material blocking, duplicate-command prevention and recovery after a network fault. Specify which supplier investigates a discrepancy between the instrument, formulation software and dispensing equipment. Require export of standards, approved formulas, spectral results and correction history in usable formats when the agreement ends. Training must include the failure cases: no candidate, uncalibrated instrument, insufficient stock, mismatched container, rejected visual match and manual fallback.

Three common questions

Can paint formulation AI start with little historical recipe data?

Yes, evaluate a conventional spectral formulation baseline and clean material master first. A learned correction layer cannot be promised to work with a tiny set of approved examples. Give every new target, candidate, trial and correction stable IDs so the evidence base grows; then test whether learning adds measurable value.

What if paint color difference passes but the customer rejects it?

Compare the viewing illuminant, gloss, texture, substrate, film thickness and possible metamerism. Check that both parties use the same physical standard and acceptance authority. Treat a mismatch in coating appearance separately from an error in pigment proportions. Do not let the software overrule a documented customer visual decision.

Should formulation software connect to the dispenser immediately?

First demonstrate repeatable candidate quality and traceability in the lab. Then connect approved instructions and actual-weight feedback. A dispenser can reproduce a bad recipe quickly if the material master is wrong. Keep a human release gate and verify that a network fault or material barcode mismatch stops the command safely.

What to prepare for the first discussion

Bring the coating family, monthly matching volume, current correction history, color standards, instrument model, approved pigments, lab and production application conditions, approval roles and equipment to connect. If some information is missing, identifying those gaps is a valid first project phase. TOMAS TECH can help map the instruments and workflow, define formulation and approval requirements, and prepare a measurable acceptance plan. Contact our team with your product family and where color matching currently stalls.

Sources and scope

• X-Rite coatings formulation workflow: vendor description of measurement, formulation, drawdown correction and release.
• X-Rite custom color matching: vendor description of instruments, software and dispensing.
• Datacolor formulation software: vendor description of spectral formulation and applications.
• Datacolor Match Pigment specification: vendor feature list.
• ASTM D2244-25: public scope of color difference and tolerance calculation; full standard is paid.
• ASTM D1729-22: public scope of visual appraisal; full standard is paid.
• ASTM E1345-24: public scope of repeat measurement; full standard is paid.