When a heat-treated lot fails its hardness specification, a team often has to reconstruct the furnace temperature history, hold time, quench conditions, material batch and inspection record. AI hardness prediction offers a way to flag a suspicious lot before its test result arrives. It does not turn a process log into a measured hardness value. The prediction is an estimate from available inputs; the hardness test is an observation made under a defined method and at a defined location.
The useful purchasing question is therefore practical: Which decision could your team make earlier if a prediction were available? For most plants, the first use case is to prioritize confirmation of high-risk lots while continuing the approved test and release process. This guide explains how a manufacturer in Thailand or elsewhere in Southeast Asia can select a pilot, join data, evaluate a model, compare vendor proposals and keep quality records intact.
Start with the decision, not the algorithm
Hardness after quenching and tempering can depend on composition, section size, heating and holding, cooling, tempering, load layout and the actual conditions experienced by the workpiece. A furnace setpoint is only one part of that story. If two loads use the same recipe but have different sizes, fixtures or quench conditions, a model that sees only the setpoint may miss the difference the process engineer already knows matters.
A useful early question is, “Which lots deserve an extra look before the scheduled measurement?” Another is, “Did a particular cooling history differ from comparable lots?” The alert may prompt an operator to verify a sensor, a process engineer to examine a curve, or quality staff to review the planned sampling position. Those actions have measurable value even when the AI never makes an acceptance decision. Using a displayed estimate alone to release an untested product is a much more demanding application and should not be assumed from a promising pilot.
State the prediction time clearly. Is the alert needed before quenching, after quenching, after tempering, or just before hardness inspection? Inputs that arrive after that time cannot be used for a genuine early prediction. Define the lot or workpiece covered by the output and the person who will act on it. This short statement is more valuable at the start of a project than a list of fashionable models.
Keep estimated and measured hardness distinct
ISO 6508-1:2023 specifies methods for Rockwell hardness testing of metallic materials. ASTM E18 describes Rockwell test procedures and cautions that a measurement at one point on a part may not represent the whole part. NIST also explains measurement error and uncertainty in Rockwell testing. These sources are a reminder that the target data themselves require controlled test conditions. A model’s deviation from a poorly identified test value is not a clean measure of model quality.
Show “AI estimate — decision support” and “Measured hardness — inspection result” as separate fields. Include the scale, such as HRC or HV, and the measurement position. Do not silently combine different scales or converted values. Preserve the input timestamp, model version and any missing input indicators. A neat number on a screen is easy to mistake for a completed test unless the interface makes the distinction explicit.
Draw a boundary around the first process
Begin with one material family, a small set of furnaces, a defined heat-treatment route and a consistent hardness test. Carburizing, nitriding, induction heating and through-hardening may call for different outputs and controls. Surface hardness and core hardness must not be collapsed into a single label. A pilot boundary might specify the steel family, equipment, recipe, inspection scale and measurement location. It should also list excluded cases: a new grade, a furnace after major repair, a changed fixture, missing logs or an overdue instrument check.
The right pilot is not automatically the most critical product. It is a process with traceable historical lots, repeatable measurements and a decision that can change when an early alert arrives. Review the product and material codes, equipment, sample method, available history and current cost of investigation. If an important line changes recipes constantly and has few comparable lots, data collection may be the first project. An “outside the validated scope” result can be safer and more useful than a forced numerical estimate.
Join process history to the actual test result
The minimum data chain links product, material lot, heat-treatment lot, equipment, treatment time and inspection. Add the data that the process engineer judges relevant: dimensions, furnace temperature curve, actual hold time, quench medium and condition, time to quench, tempering route, load quantity and position. Keep both setpoints and measured values where available. Record the test scale, test position, initial result, repeat result and reason for retest. If a furnace log cannot be joined to a paper inspection sheet by lot, the extra sensor data will not yet support a trustworthy model.
Build a data dictionary with four questions for each field: What does it mean? When is it known? Where does it come from? What happens when it is missing? Quality engineers should define the test labels; process engineers should define the process variables; IT should verify IDs and clocks across MES, furnace logs and inspection systems. Do not average away inspection positions before asking what the product specification requires. When the inputs describe a lot but the target is a workpiece or an indentation, document that mismatch and choose the modeling unit deliberately.
Clock alignment deserves a separate check. A furnace controller, MES and tester can show different time settings. A seven-minute offset could join the wrong segment of a continuous log. Preserve treatment start and end, quench start, test time and timezone. For cross-border reporting, separate stored UTC timestamps from the local time shown to a shift supervisor. Select several historical lots and trace them manually end to end before automating the join.
Beware of data leakage, meaning information that would not exist at the prediction time enters the model. Final disposition, a retest outcome or a downstream failure label can make retrospective accuracy look excellent while being unavailable at the point of use. Research on machine learning for heat-treated alloys explicitly discusses leakage from correlated mechanical properties. Freeze the prediction point and permit only data actually available then. Group and time-based validation should also keep related material batches and nearby production runs from making the evaluation too easy.
The cited 2026 heat-treated-alloy paper is a methods study using synthetic data, not evidence that a particular factory achieved an operational improvement. Use its leakage warning as a validation principle, then test any proposed system on your own traceable production records.
Design a pilot that quality staff can judge
Run the model in parallel with the existing inspection process. Split historical lots in production order: earlier data for development and later data for evaluation. Then observe new lots prospectively. Consider whether the evaluation covers maintenance, seasonal conditions and normal changes of load. If it does not, state the limit; do not market one short result as a plant-wide guarantee.
Compare AI with simple alternatives: the current control limits, a material-specific regression and a recent-lot baseline. For a numeric estimate, examine a suitable error metric, but also break out errors by material, furnace, scale, test position and proximity to the specification boundary. For an alert, count inspected nonconforming lots the model failed to flag, as well as extra checks triggered on lots that were ultimately acceptable. A high overall accuracy can be meaningless when problematic lots are rare. Conversely, flagging every lot may prevent misses while overwhelming the shift.
Agree on the alert threshold with quality staff, considering product risk and inspection capacity. Keep a pilot record for every prediction: its time, input completeness, model version, estimated value or risk tier, action taken and eventual test outcome. Investigate wrong predictions individually. A process change, inconsistent test location or missing quench measurement may explain more than an algorithm tweak. An explanation such as “cooling curve differs from similar lots” should direct an engineer to evidence; it does not prove a root cause.
Before the pilot, write acceptance criteria in operational language. For example: the share of eligible lots receiving a prediction, the share correctly marked “outside scope” when inputs fail, the number of problematic lots flagged before inspection, the number of avoidable checks and the time from alert to action. Set target values locally; no universal threshold fits every product. At the end, choose explicitly among scale up, narrow the scope and retest, fix the records first, or stop. That decision prevents an endless proof of concept with no plant owner.
Protect CQI-9 and customer-facing records
For automotive heat-treatment work, identify the applicable customer requirements and assessment framework. AIAG describes CQI-9 as a heat-treat system assessment used to evaluate and improve process control. The specific records, methods and retention periods that apply must be checked against the relevant edition, customer requirements and process. This article does not assign a universal obligation to every plant.
Keep the authoritative process log, equipment checks, calibration, work instruction, lot trace, measured hardness, retests, nonconformance disposition and approvals in their designated systems. Attach the AI output as supporting information. When an auditor or customer asks why a lot was accepted, the answer should lead to the specification, actual test, release decision and approver. A model score alone is not a complete inspection record.
Manage the model as a controlled asset. Record its training period, eligible materials and furnaces, feature definitions, validation results, release date, approval and reason for each change. Save the model version and inputs that produced each operational output. Otherwise a later retraining may change the displayed value and obscure what the team actually saw at the time. These are sound operational controls for AI; they are not a substitute for reading the applicable CQI-9 requirements.
If AI says “normal” but a test fails, follow the existing containment and quality procedure while checking the measurement and input history. If AI flags risk but the test passes, do not simply delete the alert. Review the test location, process change and alert threshold. Keep model retraining separate from lot disposition. A threshold changed automatically without quality approval could silently change the plant’s decision practice.
Compare the full cost of introduction
Separate the proposal into data acquisition, lot and inspection linkage, model development and validation, operator interface and notifications, and ongoing monitoring and change control. A plant with sensors already in place may still need substantial work to connect logs to inspected lots. A paper-heavy plant may first need a stable data-capture process. No credible fixed price follows from the phrase “AI hardness prediction” alone; the cost depends on equipment, data condition and integration scope.
Ask each vendor seven questions. Is the output for a lot, a part or a test point? When is it available? What happens when data is missing? How does performance vary by material and furnace? Can the plant export inputs, versions and decisions? Who helps investigate disagreement with testing? What do furnace additions, sensor replacements, recipe changes and extra languages cost after launch? Put the answers in one comparison table. An attractive average error and a low initial price are insufficient if the model cannot identify an out-of-scope condition or support a quality investigation.
Measure value against a baseline of retests, extra inspections, investigation time, rework and scrap. Record both the cost of false alerts and the benefit of earlier detection. If notifications become so frequent that nobody acts, predicted risk has little operational value. Scope a pilot separately from a production contract. The latter needs access controls, backups, outage behavior, update approval, support coverage and data-export terms.
Make it usable on a Thai production floor
A Thai plant may have furnaces from several suppliers, local-language operator entries and reporting to headquarters or overseas customers. Standardize part, material, furnace, shift, time and test-location identifiers before translating a dashboard. The same lot must mean the same thing in Thai, English and Japanese views. Preserve the original record, timestamp and approver for audit; present the next action in the language used by the person on shift.
Design for network outages. Testing and process logging continue even if a prediction service is unreachable. The interface should distinguish “not predicted,” “missing input,” and “outside validated scope” from “normal.” After maintenance or sensor replacement, check whether input distributions have changed before trusting the previous validation. Link maintenance events to model monitoring. In daily use, review the previous shift’s test outcomes and unresolved alerts, confirm who receives a new alert, and reconcile each prediction with the later measurement. This should fit the existing shift and quality report rather than require a disconnected AI workflow.
Training should include failure cases. A normal prediction does not cancel scheduled testing. “Outside scope” does not mean pass. A failed test follows the quality procedure even if the model predicted normal. Distinguish a hardness risk alert from an equipment safety alarm; they do not have the same meaning or response. Night and weekend coverage also matters: an alert delivered when nobody can act may not shorten the decision time.
Five common failures and how to respond
Only the furnace setpoint is available. The actual load, workpiece position, fixture and quench may differ between runs. Add measured process history and operator conditions where feasible; if key factors cannot be recorded, narrow the validated scope. Buy another sensor only after identifying which missing variable blocks prediction or investigation.
Different hardness results are collapsed into one field. HRC and HV, surface and core, initial and repeat tests are different targets. Keep scale and location, preserve the initial result and reason for retest, and use separate targets or models when appropriate. The data dictionary is part of model design.
A retrospective score drops in production. Check whether adjacent lots or a common material batch appeared in both training and evaluation, and whether an input was known only after the intended prediction time. Then check maintenance and recipe changes. Random historical splits do not demonstrate future-lot performance; use a time-based evaluation and monitor deployment.
The alert queue overwhelms the shift. A very conservative threshold can flag almost every lot. Show the reason, lot, priority and required check for each alert, and measure whether the team can respond. Any threshold change that increases missed cases needs quality review and a documented evaluation.
Nobody can explain a model change. Furnace repair, new suppliers and recipe adjustments can change the input distribution. Compare the old and new versions on the same validation set, record what changed and who approved it, and keep a rollback path. Without version history, a disagreement between prediction and test is harder to investigate.
A practical implementation sequence
- Define one use case. Name the line, material family, hardness scale, specification, current test practice and earliest useful prediction time.
- Trace sample lots manually. Walk from furnace curve through lot ID to inspection sheet, including retests and changes. Fix broken identification before model work.
- Freeze a validation plan. Specify inputs available at the prediction point, exclusion cases, time-based evaluation, baselines and quality-approved metrics.
- Run alongside inspection. Display predictions without changing release decisions. Record alerts, actions, measured outcomes and missed cases.
- Approve production controls. Define the owner, versioning, outage state, retraining approval and rollback. Validate each newly added material, furnace or recipe.
Common questions
Can temperature alone predict hardness?
Temperature matters, but composition, dimensions, quench, tempering, fixture and test conditions can also matter. A simple temperature-only baseline may work within a very narrow route; it is not evidence that one sensor generalizes to every part. Compare it with a broader model and the actual test results.
Can AI reduce hardness testing?
Do not assume so. Check customer specifications, the control plan, applicable test methods and approval route before changing testing. The first practical use is usually to prioritize checks and speed investigation while maintaining the approved measurement process.
How many lots are required?
There is no universal number. The answer depends on process stability, material and furnace variety, frequency of unusual outcomes and acceptable miss rate. Coverage of maintenance, recipe changes and specification boundaries matters as much as raw count. If evidence is thin, narrow the scope and continue collecting consistent records.
Can an AI vendor take responsibility for CQI-9 compliance?
The vendor can support data collection, prediction and change history, but the plant must identify the applicable customer requirements and approve its process and quality decisions. Define the vendor’s evidence and support duties in the proposal. Have the quality team review the requirements and acceptance criteria.
Can an older furnace be included?
Sometimes. If a reliable log can be joined to a lot and its measured result, a pilot may begin without new sensors. Where logs are paper-based or IDs do not align, record capture and traceability come first. Upgrade decisions should follow a specific missing-data problem.
Begin with one traceable lot
AI hardness prediction is most valuable when it helps people identify a lot needing attention before measurement is complete. Choose a bounded route, link process and test history by lot, compare against simple baselines and evaluate misses as well as workload. Preserve measured hardness and the authoritative quality record. For related reading, see our heat-treatment and CQI-9 process-record guide and quality data management guide.
TOMAS TECH helps manufacturers connect equipment data, MES/ERP and inspection history. Share one example lot, its furnace log and inspection record to identify a sensible pilot and the data work needed first. Discuss a heat-treatment data and AI pilot.
