Implementing AI for Paint Film Thickness Prediction: Measurement, RFP, FAT and SAT

Installing a painting robot does not by itself make coating thickness consistent. Even if the robot repeats the same path, workpiece orientation, gun-to-surface distance, travel speed, paint viscosity and flow, electrostatic settings and booth airflow may change the film deposited on each surface. The purpose of a film thickness prediction system is not an attractive dashboard. It is to identify which measured dry-film-thickness location on which part may move outside specification, early enough for a named person to check a defined process condition. This guide builds a practical project for a Thai paint line, from measurement and data structure through prediction, operator action, RFP, factory acceptance testing (FAT) and site acceptance testing (SAT).

For robot types, electrostatic versus powder coating, and investment calculations, see our painting robot implementation guide. Here we focus on using actual film-thickness measurements to find process drift on a new or existing line. Equipment selection for hazardous areas also needs its own review; see the explosion-proof collaborative robot guide. An AI project does not replace that safety work.

Define exactly what the model must predict

“Predict paint quality” is too broad to procure or test. Appearance, colour difference, gloss, adhesion, corrosion performance and thickness are different quality characteristics. For this project, define the target as measured dry film thickness at specified locations and coating layers, assessed against the applicable drawing and customer specification. A part-level average can hide thin edges and heavy buildup in recesses. A single reading cannot automatically stand for the entire part.

ISO 2808:2019 describes methods for measuring wet-film thickness, dry-film thickness and uncured powder-layer thickness. Choose the production stage that the quality specification governs and put its method and location in the plant inspection standard. The gauge manufacturer DeFelsko explains calibration, verification and adjustment, which can inform a plant-specific pre-use check and substrate-appropriate setup. Its coating-thickness measurement guide distinguishes wet- and dry-film methods. Do not allow one reading to stand for the entire surface without a defined sampling plan. Drawings, substrate, roughness, coating layers and customer acceptance rules determine that plan.

Specify when the prediction is made. Is it before painting from the part geometry and recipe, during painting as speed and flow arrive, or before inspection after curing? Each point in time has different available inputs and response requirements. A model tested with inspection results that arrive after spraying may score well in a spreadsheet yet be unusable on the line. State the prediction event in the RFP and allow as inputs only signals that are reliably available at that event.

Separate prediction, disposition and control

Start with an advisory mode: show the predicted value, its uncertainty and its distance from the upper and lower specification limits. Quality can request extra measurements; maintenance can inspect the pump, flowmeter, atomizer or booth. A later stage may recommend a recipe within an approved envelope. Writing AI output directly to the robot, pump or electrostatic power supply is closed-loop control and requires a separate safety and quality change process, hard limits, interlocks, authority, fallback recipe and revalidation. Prediction accuracy alone is not approval to control an actuator.

Make measured thickness trustworthy ground truth

A training row needs more than a part number, date and thickness. Join part ID, production order, fixture orientation, coating layer, face and measurement-position ID, reading time, gauge ID, inspector, cure conditions and rework status. For multilayer coating, state whether a reading covers one layer or the total film. If a part goes through the booth twice, the result of the second pass must not be attached to the settings of the first pass.

Inspection and process events must identify the same physical workpiece. If the workpiece is scanned only at the booth entrance, define how conveyor changes and rejected parts are tracked. First-off pieces after colour change, material-lot change and cleaning may form different distributions; record those events rather than silently mixing them into “normal production.” If two robots coat one part, preserve robot number and pass sequence.

Validate the measurement system as well as the model. Record calibration, pre-use verification and substrate-appropriate adjustment. Check the effect of standard foils versus real parts and agreement between operators. Curved edges, rough surfaces and probe contact can affect readings. Compare methods before pooling different substrates in one model. Reducing model error against unstable labels does not make acceptance against a customer drawing stable.

Sampling creates selection bias. If operators measure mainly suspicious parts, the recorded defect rate is not the production defect rate. Include a defined sample of uninterrupted normal production and distinguish it from extra measurements after alarms. Give every measurement point a meaningful position ID such as edge, recess or appearance face, and version the map when the drawing changes. Moving a point without changing its ID destroys comparability.

Figure 1: Join paint settings and the thickness reading by part ID and measurement location. The ground truth is the reading made with a defined procedure, not the predicted number displayed on a screen.

Capture how the part was painted, not just machine tags

The first input family is robot operation: program and recipe ID, path segment, actual speed, gun or bell distance and angle, trigger state, pass overlap and robot assignment. Keep planned CAD values separate from actual logs. ABB’s painting software description connects spray parameters, path speed and distance with uniform coating. Those are defensible candidate features; which ones matter on your line still requires measurement.

The second family is material and applicator: paint SKU and lot, mixing recipe and ratio, viscosity, temperature, solids, measured flow, pump pressure, trigger timing, bell speed, shaping air and electrostatic voltage or current where relevant. Each machine exposes different signals; an electrostatic value is not a universal requirement. ABB’s atomizer material discusses airflow, paint properties, atomizer condition, robot path and part geometry as influences on deposition. Its product-specific performance figures are not a guarantee for another installation. In a Graco vendor case, recipe, mixture, flow and robot speed are connected to a specified film; it is an example of how to describe a process, not a universal target for your plant.

The third family is the part and the environment: geometry, curvature, recesses and edges, fixture position, pretreatment, substrate, booth temperature and humidity, ventilation state, filter differential pressure, conveyor speed and cure profile. All need not enter the first model. But a large unmeasured change may make the model blame robot speed for a material or airflow problem. Correlation is not proof of cause. Process engineers must compare model suggestions with maintenance and change records.

Select inputs by meaning and reliability, not simply by whether a tag exists. Was the value available before prediction? Is it captured for every part? Is its unit stable? Was the sensor healthy? Did its meaning change with a recipe revision? A flow setpoint is not the same as measured delivered flow. Maintain a data dictionary with tag names, units, update rate, time zone, missing-value behaviour and owner.

Align events across systems

Spray settings can change by the second, while inspection may occur after cooling. Unsynchronised PLC, robot, MES and gauge clocks can attach the wrong pass to the right part. Store entrance, spray-start, spray-end and inspection as separate events. If late data updates an earlier prediction, retain the earlier version so the action actually presented to the operator can be reconstructed.

Compare the model against three baselines

Before inviting AI proposals, define comparators: a historical mean by part and coating layer, an engineer’s warning threshold, and the current inspection and adjustment routine. A sophisticated model that slightly improves numerical error but creates a large inspection burden has little operational value. A model that helps find a risky location before inspection can be valuable even if its numeric error is imperfect.

Use two levels of metrics. For regression, inspect absolute error and its distribution by product, colour, material lot, machine and measurement location. For decision support, inspect missed lower-limit and upper-limit violations separately, false alerts, added inspection, response time and rework. Do not merge wide and narrow tolerances into one flattering average. Contract targets should follow the customer’s tolerance, gauge uncertainty and cost of extra inspection. No public source establishes a universal “90% AI accuracy” for every paint booth.

Avoid a random split that places parts from the same paint lot and shift in both training and testing. Their shared paint and environment can overstate future performance. Hold out later production periods and examine new colours, products, lots, seasons and booth conditions separately. Specify when the system must say “outside model scope” rather than predict an unfamiliar product. With small datasets, a calibrated regression and process rules may be easier to maintain than a complex model.

NIST’s manufacturing monitoring research illustrates research on AI-enhanced process monitoring; it does not report the benefit of this proposed thickness application. Once deployed, monitor both input distribution and prediction errors. Material changes, a new bell, a fixture or program revision can call for revalidation. AWS machine-learning operations guidance likewise describes monitoring data and model-quality drift and governing retraining.

Figure 2: Path, flow and booth signals feed a prediction by measurement position; measured thickness closes the learning loop. In advisory mode, no settings change without an authorised review.

Make the operator display actionable

A heat map alone does not tell the shift what to do. Show part ID, product, colour, layer, position map, predicted thickness, both limits, missing or unreliable inputs and recent measured results together. Connect alerts to approved instructions: add readings at positions A and C, check the flowmeter, or hold the first part after a colour change. The colour of a heat map must not silently become final product disposition.

Version the model, recipe, measurement map, input cut-off and alarm threshold. Otherwise an auditor cannot reconstruct why a part was held. Record an operator override and its reason, but keep everyday interaction light enough to use. MES or QMS should remain the system of record for quality hold and release under existing authority; avoid a second, conflicting disposition state on the AI screen.

Set thresholds with measurement uncertainty and workload in view. During a trial, report alerts, confirmed out-of-limit readings, misses, inspection time and line stops in one table. Treat upper-limit risk separately from thin film because overbuild can affect downstream operations as well as material usage. The customer’s specification and the plant’s inspection procedure remain the final acceptance basis.

Stage a 60–90 day validation project

The duration is an illustrative project window, not a fixed rule for all lines. First define positions and a data dictionary, then record the current thickness distribution, rework, inspection effort and paint use. Choose one coating layer, representative parts and one operating booth. If historical data are sparse, stabilise capture before commissioning model development.

Next connect each workpiece to conditions and inspection. Review missing data, clock offsets, gauge checks and rework identification every week. During modelling, compare baselines, test on a later time period, review outliers and discuss false alerts with operators. Finish with a shadow trial: record advice without changing process settings, measure response time and compare predictions with actual readings. Use the result for a go, expand or stop decision.

Budget for more than software: gauges and verification, part identification, robot and PLC logs, network, storage, MES/QMS interfaces, sensor repair, displays, training, support, retraining and validation of new products. Compare benefits at like-for-like throughput, product mix and quality requirements. Track rework labour, added inspection, paint consumption, scrap and customer complaints. If cleaning or a recipe change produced part of the improvement, do not double count it as a model benefit.

Put measurable requirements into the RFP

“Optimise thickness with AI” is not a testable request. Give every bidder the same scope and state whether any unknown value is to be set by the plant or decided in a paid discovery phase.

Clarify whether the supplier owns data acquisition, gauge integration, operator display, support and retraining. Separate authority over painting recipes, gauge management and model updates. Require review and rollback for remote model changes. If a generative-AI explanation is added, retain its underlying measurements and model output as structured records; prose alone cannot be the quality record.

FAT: test exceptions, not just the dashboard

At FAT, run representative test data through part identification, time alignment, units, position maps, model version, alerts, roles, logs and backup. Include missing and late signals, a wrong lot, an unverified gauge, a recipe switch, a repainted part and an unseen product. For each case predefine whether the system predicts, reports “outside scope,” requests measurement or escalates to a person. Silent substitution of an average can conceal a thin-film risk.

FAT performance on prepared data does not prove performance in the actual booth. List products and periods represented in the test set, and carry unresolved signal availability to SAT. Quality, process engineering, maintenance, IT and the shift lead should agree on acceptance criteria.

SAT: verify the line and its response procedure

At SAT, trace real parts through scanning, robot and PLC logs, gauge readings and alert delivery. Include continuous running, colour change, start-up, stop and recovery, material-lot change, communication loss and restoration. Assess numerical error alongside out-of-scope behaviour, additional inspection, hold release, audit logs and operator understanding. If the AI fails, the line must be able to follow its approved recipe and inspection procedure or stop under defined conditions.

The point of FAT and SAT is also to agree what the plant does when a reading is impossible, data arrive late, or the responsible person is absent. Record open items, temporary rules, owners, due dates and retest criteria. Define the period of performance support and the material, machine or product changes that trigger revalidation.

Figure 3: Carry measurement requirements from the RFP into simulated FAT exceptions and actual SAT line events before approving advisory operation.

Common failure modes

Assuming an image reveals film thickness. A camera can help with appearance and position, but visible colour variation is not the same as measured dry-film thickness. Verify whether a specified optical instrument directly measures thickness; otherwise retain a calibrated gauge reading as ground truth.

Losing part identity. A daily average transcribed from paper cannot be aligned with robot passes and material lots. Start by identifying the workpiece, face, layer and measurement position even if some collection remains manual.

Training only on easy good parts. This leaves no evidence of error near the limits. Include historical deviations, start-ups and colour changes alongside a regular normal-production sample. There is no need to create unsafe conditions deliberately.

Auto-correcting because predictions look good. Association does not establish the effect of changing flow or voltage on thickness, appearance and equipment safety. Begin with authorised advice; use formal change control for closed-loop operation.

Freezing the model. Paint formulation, bell, filters, fixtures, robot paths and curing can change the relationship between inputs and thickness. Document monitoring, update criteria, owner and rollback.

Eight questions before requesting a quotation

1. Which layer, products, faces and positions are in scope, and who owns the upper and lower limits?
2. Which gauge and procedure measure dry thickness, and how are calibration, verification and adjustment recorded?
3. Can part ID, spray pass, paint lot and measured location be joined without guessing?
4. When is a prediction displayed, and which signals reliably arrive before then?
5. How will the cost of missed low or high film be weighed against extra inspection?
6. What appears when data are missing, a product is unknown, or a sensor fails?
7. Who approves advice, recipe changes and quality hold or release?
8. Which records decide FAT and SAT acceptance, and who pays for retesting?

Answering these questions lets you compare measurement quality, data connection and operational responsibility alongside price. If a key answer is unknown, procure a short site survey and measurement design first. Our broader AI process improvement guide covers project selection and benefits; this article applies that approach to a paint booth and specific measurement locations.

Define scope for high-mix, low-volume lines

Many Thai component plants run several products, colours, layers and fixtures through one booth. Promising one model for the entire line can create predictions for combinations with too few measured examples. Group parts by surface geometry, required film thickness, paint chemistry, robot path and measurement map rather than by part-number count alone. A flat panel and a deep enclosure recess can deposit paint differently even when they share a colour. Parts that differ mainly in size may be more comparable when their faces and paths are similar. Confirm either assumption with measured data before defining the model’s scope.

Keep three lists: products predicted now, products to add after data collection, and products staying on the current inspection procedure. On the screen, give an out-of-scope message and the applicable inspection instruction instead of a fabricated estimate. When adding a colour or product, review the measurement map, customer limits, recipe, normal and exceptional samples, false alerts and misses. Inputs lying within historical numerical ranges do not prove that their relation to thickness is unchanged. Quality should approve reuse of the model.

Choose a representative part and measurement plan

The first part should not be selected solely because its production volume is high. It should offer both surfaces where thickness problems matter and geometry that inspectors can measure repeatedly. Measuring only a flat appearance face can hide weaknesses when the model is extended to complex workpieces. Starting with the deepest recess, on the other hand, can make gauge placement unstable and obscure whether error comes from the model or the measurement. Select a small set of contrasting locations such as an appearance face, an edge and a recess, and establish repeatability at each one. Preserve any customer-mandated points; distinguish process-improvement points in the record.

Separate scheduled samples during continuous production from event samples after colour change, start-up, cleaning or material-lot change. Both categories must join to the process log by part ID. If an extra trial measurement influences production disposition, define its relationship to the formal inspection procedure and who approves it. State who verifies an imported reading and whether correcting a transcription error triggers recalculation of model evaluation. Collecting a large dataset without these rules can create a costly label-reconstruction exercise later.

Count the different costs of wrong predictions

Missing a thin-film defect and alerting on a good part have different costs. The first can lead to additional inspection, rework, in-process hold and customer escape risk. The second consumes inspector time, can delay the line and may prompt unnecessary adjustments. Missing excessive thickness also has implications beyond paint cost, depending on assembly and appearance requirements. Do not compress these into one “accuracy” percentage; assess them by surface and customer requirement.

During the trial, measure a defined sample of parts that received no alert as well as the alerted parts; otherwise missed defects remain invisible. When an alert says “flow dropped,” check whether a sensor fault or a fixture change occurred at that moment. Connect the operator’s response to the actual reading and investigate unhelpful alerts. Moving a threshold to reduce alerts may increase misses, so compare before and after under a similar product mix and sample plan. Version the threshold and retain the approver.

At the weekly review, examine representative wrong predictions alongside the physical part or image, process log and gauge record. Decide whether to revise the model, repair a sensor or stabilise the measurement procedure. If the cause is unknown, assign an owner and a test for the next week rather than filing every event as “poor AI accuracy.” These lessons can improve coating conditions and inspection standards as well as future model training.

Contract for data handling and plant ownership

If a supplier develops the model in a cloud environment, paint formulations, part numbers, process settings and inspection records may contain confidential information. Specify storage location, authorised users, cross-border transfer if any, retention, deletion and return at contract end. State whether the supplier may reuse the model or processed data for other customers. Decide cloud deployment under the plant’s IT policy and contract. If connectivity is intermittent, define how painting and inspection continue when the prediction service is unavailable.

Day-to-day ownership also stays at the plant. Quality owns gauges and specifications; production owns recipes and work instructions; maintenance owns sensors and equipment; IT owns connections and access; the AI owner monitors model performance. Define who checks the physical part, places a hold and releases it, including nights and holidays. Monthly review should cover missing data, out-of-scope products, alert response, rework, manual overrides and model revisions as well as numerical error. Only then can prediction lead to durable process improvement.

Conclusion

An AI film-thickness project begins with the definition and reliability of measured dry film thickness. Align part, face, layer, location and time; connect robot path, speed, flow, electrostatic settings, material and environment to the same production event; and compare the model with simple baselines. Then create an alert that an operator can act on, and test exceptions through RFP, FAT and SAT. Prove value and safety in advisory mode before considering any control extension.

TOMAS TECH can help Thai factories define measurement locations, determine which existing machine data can be captured and scope a realistic pilot and acceptance test. Bring a representative part, inspection standard, machine list and the locations where thickness drifts. Contact TOMAS TECH to discuss the first site review.

References

• ISO 2808:2019 — Determination of film thickness
• DeFelsko — Calibration, verification, adjustment
• DeFelsko — Coating thickness measurement
• ABB — Painting Software Solutions
• ABB — RB 1000i-S atomizer
• Graco — Different colours and recipes in a paint installation (vendor case)
• NIST — AI-enhanced manufacturing monitoring
• AWS — Automated retraining framework