AI process parameter recommendations are suggestions from a trained model about which machine settings to change, and by how much, to keep quality, energy use or scrap on target. The model does not run the line. It reads your process data, compares it with past batches and tells the operator or process engineer what to adjust. This guide explains how that works, which data you need, how to test the recommendations before they reach the shop floor, and which Slovenian and EU rules to keep in view.

TL;DR

• AI recommendations use pre-trained machine learning models and explainable AI to suggest parameter changes that may prevent failures or quality degradation.
• Start with one target variable, such as first-pass yield or scrap, then select only the parameters, context and outcomes that relate to it.
• Clean, aligned and labelled data matters more than volume. Plant specialists must decide what counts as a real defect.
• Test every recommendation offline and in shadow mode first. Keep a human approving changes.
• Begin with a narrow pilot on one line. EU AI Act obligations for AI embedded in regulated products apply later than many managers expect.

Inhaltsverzeichnis

• What AI parameter recommendations are
• Choosing the parameters and data
• Preparing equipment and production data
• Testing and validating recommendations before production use
• Applying recommendations safely on the shop floor
• EU and Slovenian rules to keep in view
• A practical pilot plan
• Our perspective
• From data to recommendations with Mestric
• FAQ
• Quellen

What AI parameter recommendations are

Every process has settings: temperatures, pressures, speeds, dwell times, feed rates, recipe nominals. Operators set them from experience, work instructions and the odd trial. Parameter optimisation, as academic work on the subject describes it, means finding the best values within defined limits to improve product quality.

An AI recommendation adds a learning layer. The model has seen thousands of past runs, with their settings, conditions and results. When a new run starts to drift, it suggests a change that, in comparable past situations, led to a better result.

That is different from a fixed alarm limit. A limit tells you something has crossed a line. A recommendation tells you which setting to move, in which direction and why.

Problems it can address

• Quality drift: gradual loss of quality as material lots, tools or ambient conditions change.
• Start-up scrap: in an injection-moulding case study, AI-assisted start-up tuning cut mean start-up scrap by 94% against a junior-operator baseline. That is one study in one setting, not a promise for your line.
• Energy waste: in Slovenia, models built on data from more than 2,000 electric arc furnace batches at SIJ Acroni predicted energy consumption and optimal melting time, and the authors expected the approach to lower consumption.
• Machine failures: settings changes that reduce stress on equipment before a breakdown.

Choosing the parameters and data

The most common mistake is to collect everything and hope. Work in this order.

1. Pick the target variable. Product quality is the usual choice. Define it in numbers: scrap rate, first-pass yield, a critical-to-quality dimension, kWh per batch.
2. Capture the starting conditions. Industrial data science guidance for batch processes lists raw-material properties, time-varying sensor trajectories across process phases, event data that identifies phase and batch, and final product properties.
3. Add context. Tool wear, material attributes, ambient temperature and humidity often explain differences that the machine settings alone cannot.
4. Record outcomes and interventions. If a recommendation was applied, store what was changed, who approved it and what happened next. Without this, the model cannot learn from its own advice.

What a useful dataset contains

A published list of what to select for parameter optimisation follows the same pattern: recipe nominals and ranges, machine tags, lot genealogy, quality outcomes and context. For recommendation use cases, one industrial-data guide adds that you should keep the human review decision, the action taken, the post-action reading and the version of the model that made the suggestion.

Context is easy to underrate. Surface finishing shows why: preparation steps, batch records and coating quality documentation all sit beside the process settings, and an AI model needs that same surrounding record to tell a settings problem from a material or environment problem.

Let domain experts decide what is relevant

Slovenia’s SRIP Factories of the Future makes a point worth repeating. In its commentary on AI in production, domain experts have to identify which variables are genuinely relevant, separate defects that block a sale from harmless marks, and agree on the correct training data. Volume alone does not do this.

In practice, put a process engineer, a quality technician and an operator in one room before any modelling. Ask them which five settings they would change first when quality slips. Start there.

Preparing equipment and production data

Raw machine data is rarely ready for a model. Typical work includes:

• Aligning time. Sensors sample at different frequencies and signals lag behind actions. Research on process data pre-processing covers handling sampling frequencies, lag times, missing values and outliers, followed by sensor fusion and feature extraction.
• Labelling machine states. A Slovenian paper on energy forecasting for production processes notes that machine states must be labelled and data frequencies harmonised. It lists sensors and meters, MES and SCADA histories, energy data, production operations, operating conditions and external factors such as weather and energy prices as inputs.
• Linking quality results to the exact run. A scrap figure at the end of the shift is not enough. Each result needs the batch and machine it came from.
• Removing or flagging bad data. Calibration faults, manual overrides and test runs distort learning.

This is where a manufacturing execution system earns its place. When machines connect directly and data is captured automatically, the batch identity, downtime reasons and quality readings are already tied together. Our article on the role of data in manufacturing covers the types of data and how to collect them, and the piece on real-time data integration with MES platforms shows how the connection works.

Check your readiness first

Rok Prešeren, speaking about when a company is ready for AI in production, points to a sequence: digitise key processes, consolidate data from machines, sensors and systems, set up basic analytics and KPIs, train people to use and trust the data, and run small pilots before scaling. If your downtime reasons are still written on paper, fix that first. The model cannot recommend what it cannot see.

Testing and validating recommendations before production use

No single page we found sets out one standard method here, so this section reflects common engineering practice rather than a cited rulebook. Adapt it to your quality system and customer requirements.

1. Offline back-testing. Run the model on historical batches it has not seen. Compare its suggestions with what the best operators actually did. Where the model and your best operators disagree, review the cases together.
2. Range checks. Reject any recommendation outside the validated process window, the equipment limits or the product specification, whatever the model predicts.
3. Shadow mode. Let the model make recommendations on live production, but do not apply them. Log what it would have suggested and compare with the results. Run it long enough to cover different materials, shifts and seasons.
4. Controlled trial. Apply recommendations on one line, product or shift, and keep a comparable baseline. Measure scrap, yield, cycle time and energy against that baseline.
5. Approval and rollback. Define who approves a change, how to revert it, and what triggers a stop.
6. Ongoing monitoring. Track how often advice is accepted, rejected and successful. Retrain when materials, tools or products change.

Explainability makes steps 1 to 3 faster. Techniques such as LIME or counterfactual generation show operators how to change input features to reach a desired quality outcome, so they can judge whether a suggestion makes physical sense. The same research stresses that human verification remains essential.

What to measure

Also watch for limitations. A model only knows the conditions it has seen. A new raw material supplier, a worn tool or an unusual season can push it outside its experience. Treat a sudden fall in acceptance rate as a signal to investigate, not a training problem for operators.

Applying recommendations safely on the shop floor

Process safety and product quality come before optimisation. A few rules keep this simple:

• Recommendations appear to the operator with a reason, such as the factors that drove them.
• Safety-related parameters are locked or outside the model’s scope.
• A person confirms any change that affects a safety, regulatory or customer-specified characteristic.
• Every change is logged with time, user and result, so quality audits can trace it.

To connect advice with live results, you need a view of KPIs, downtime and quality in one place. Our guides on real-time performance tracking und quality monitoring workflow describe how to set that up.

EU and Slovenian rules to keep in view

Most parameter recommendation systems are decision support for engineers, not consumer-facing AI. Even so, three recent developments are worth tracking, and your legal or compliance adviser should confirm how they apply to your system.

• National implementation. The Slovenian government adopted a proposal for a law implementing the EU AI Act, assigning supervision of high-risk AI systems to ministries and bodies including AKOS.
• Transparency duties. Certain Article 50 transparency obligations began to apply in Slovenia on 2 August 2026.
• Products covered by existing EU legislation. Under the Digital Omnibus Regulation (EU) 2026/1744, rules for high-risk systems integrated into such products apply from 2 August 2028.

If your AI sits inside a machine or a regulated product, ask your supplier which category it falls into. If it only advises your own staff on settings, keep documentation anyway: data sources, model versions, test results and approvals. You will need them for quality audits, whatever the AI rules say.

A practical pilot plan

A narrow pilot is better than a plant-wide rollout, which matches the staged approach Prešeren describes above.

1. Weeks 1 to 2: choose one line and one target, such as scrap on a single product family. Record the current baseline.
2. Weeks 3 to 6: connect machines, fix data gaps, label states and link quality results to batches.
3. Weeks 7 to 10: build and back-test the model with process experts reviewing disagreements.
4. Weeks 11 to 16: run in shadow mode, then a controlled trial with operator approval.
5. After the trial: compare with the baseline, decide whether to extend, and write down what you learned.

The timings are a planning example, not a benchmark. Your data situation sets the real pace. For a wider view, read our step-by-step production optimisation guide and the article on optimising production workflow with AI.

Our perspective

We see the same pattern in factories again and again. The model is rarely the hard part. The hard part is that the data is split between machines, spreadsheets and people’s memories, and nobody can say which batch produced which result.

So we start with the foundation. Connect the equipment, capture data automatically, tie it to orders and batches, and watch the KPIs for a while. Only then does an AI recommendation have something solid to learn from. Operators trust advice that matches what they see on their own line, and they stop trusting it quickly when it does not.

We also believe the recommendation should be easy to read. A plant manager or supervisor should see the suggested change, the reason and the live result in one screen, without an analyst in between. That is how we built Mestric: simple to use, connected to the machine, and focused on the production manager’s decisions.

— Andraž

From data to recommendations with Mestric

Mestric™ is a Manufacturing Execution System that connects directly to your equipment and collects performance, downtime, quality and cost data in real time. It adds AI-powered optimisation to help you find bottlenecks, reduce manual errors and improve process norms. You can see it running in a real production environment through an on-site presentation. Learn more about Mestric™ MES, or see how it compares with traditional manufacturing approaches.

FAQ

Do we need a data scientist to start?

Not necessarily for a first pilot. You do need someone who knows the process and can define what a defect is, plus reliable machine data. Experts on your own floor decide which variables matter, which is more important than modelling skill at this stage.

How much data do we need?

There is no universal number. Quality and relevance count more than volume, as the Slovenian SRIP Factories of the Future stresses. The SIJ Acroni study used more than 2,000 furnace batches, but a simpler case may need fewer, and a complex one more.

Can the AI change machine settings automatically?

Technically it can, but we recommend starting with advice that an operator approves. Human verification is considered essential in the research on explainable recommendations. Move towards automation only after shadow mode and a controlled trial show stable results.

Does the EU AI Act apply to our factory?

It depends on what the system does and whether it is built into a regulated product. Some obligations already apply in Slovenia, and others for products covered by existing EU legislation apply from 2 August 2028. Ask your legal adviser to classify your specific use case.

Quellen

• MOXAI: Manufacturing Optimization through Model-Agnostic Explainable AI
• Manufacturing Process Parameter Optimization
• Industrial Data Science for Batch Manufacturing Processes
• Napovedovanje energetske porabe proizvodnih procesov
• SRIP Factories of the Future: AI in production, data quality
• Process Parameter Optimization
• Data Pre-processing and Sensor-Fusion for Multivariate Processes
• Umetna inteligenca v proizvodnji: kdaj je podjetje zrelo
• Data Readiness for Industrial AI Agents
• Data-Driven Modelling and Optimization of Energy Consumption in EAF
• Capturing mold qualification sensitivities for adaptive injection moulding
• Slovenija med vodilnimi pri uvajanju evropske uredbe o umetni inteligenci
• Začel veljati Digitalni omnibus o umetni inteligenci
• Začela so se uporabljati pravila o preglednosti sistemov umetne inteligence
• Industrial coating capabilities and quality documentation

Empfohlen

• Leitfaden zur Verbesserung des Fertigungsprozesses für Effizienz
• Role Of AI In Manufacturing – Impact On Efficiency
• Manufacturing Optimization Checklist 2026: Cut Costs 20%
• How to optimise production workflow with AI in 2026