Manufacturing process parameter optimisation

Quality and cost of many products depend on dozens of interacting machine settings that operators tune by experience and trial and error. AI models learn the relationship between settings, material state and outcome from historian and quality data, then recommend (or, under strict safety limits, directly apply) better set-points. Typical goals are less scrap, higher throughput and lower energy per unit, with the autonomy level rising step by step from advice to closed loop.

Problem & value

• Today: parameter windows come from datasheets, DoE campaigns and the experience of senior operators. Settings are rarely re-tuned when raw material, weather or machine wear drift, so plants run conservatively. Know-how leaves with retiring staff.
• AI value: learn non-linear, multi-variable interactions that single-loop PID and fixed recipes miss; keep settings near the optimum as conditions drift; shorten ramp-up of new products with fewer experiments (Bayesian optimisation); capture expert knowledge in models.
• Benefits: vendors report savings in energy, raw material and off-spec product; these figures are site-specific and vendor-published, so measure your own baseline before committing to a business case [3][4]. Do not rely on generic percentages.

Variants

• Advisory (level 1): model suggests parameter windows; operator decides and enters them.
• Supervised / semi-automatic (level 2): system proposes set-points, operator confirms with one click; every change is logged.
• Closed-loop AI-based APC (level 3): reinforcement learning or model-predictive agent writes set-points to the DCS/PLC autonomously inside a hard safety envelope. Documented in chemical production (Yokogawa's FKDPP algorithm in a plant of ENEOS Materials/JSR) and in process-industry products such as Imubit's closed-loop AI optimisation [1][2]. Continuous plants with a stable historian are covered in more depth in OPT-038.
• Discrete manufacturing: injection moulding, extrusion, welding, CNC, coating, heat treatment, battery electrode coating; per-cycle parameter tuning plus lot-to-lot compensation.
• New-product ramp-up: sample-efficient Bayesian optimisation / active learning over a DoE.
• Process industry recipes and blends: see OPT-031. Additive manufacturing in-situ control: VIS-057.

How it works

1. sensors/PLC/DCS/MES
2. historian (OPC UA, MQTT)
3. data prep (batch/cycle alignment, lab-result join)
4. process model (gradient boosting / neural net / hybrid physics-ML / digital twin)
5. optimiser (Bayesian optimisation, MPC, RL, MIP/NLP solver) under constraints
6. safety layer (hard limits, rate-of-change limits, plausibility check, fallback to baseline control)
7. operator review (advisory) OR write-back via OPC UA to controller set-point (closed loop)
8. outcome measurement
9. model retraining / drift monitoring

• Process model: a surrogate predicting quality and energy from settings and disturbances (material properties, ambient, wear). Hybrid first-principles + ML models extrapolate more safely than pure black boxes.
• Optimiser: Bayesian optimisation for expensive experiments; model-predictive control for continuous processes; RL (trained offline on historian data and simulators, then fine-tuned under supervision) for dynamics too complex for classical MPC.
• Safety envelope: the AI never talks to actuators directly. It only moves set-points of existing base controllers, within limits enforced outside the AI (in the DCS/PLC or a safety-rated layer), with a watchdog that reverts to conventional control on anomalies, loss of data or operator override.
• Human-in-the-loop: process engineers define objectives and constraints and approve model releases; operators can disable the AI at any time; changes to recipes under quality regimes (GMP, IATF 16949) go through change control.
• LLM role: natural-language explanation of recommendations, shift-log summaries, and translating engineer intent into solver models (see OPT-018). LLMs do not make control decisions.

Data preparation

• Which data: historian time series (PLC, SCADA, DCS via OPC UA), MES batch and recipe records, lab and inline quality results, material batch data, maintenance logs.
• Where it lives in an SMB: machine controllers with short local buffers, a SCADA or MES database, lab results in Excel or a LIMS, recipes on file shares, shift notes on paper.
• Typical gaps: inconsistent tag names, too little variation in settings to learn cause and effect, lab results that arrive days late and cannot be joined to a cycle, sensor faults, undocumented recipe changes, paper logbooks.
• Steps: (1) inventory tags and sources, see PREP-001 and PREP-005 for exports; (2) start continuous logging to a time-series store (TimescaleDB, InfluxDB); (3) align cycles and batches and join lab results; (4) clean outliers and flag sensor faults with the process engineers; (5) plan step tests or a DoE, because observational data alone rarely identifies cause and effect; (6) digitise paper logs, see PREP-008.
• Share of effort: typically 50-70 percent of the project. Check readiness first with GUIDE-003.

Implementation paths

Which path for whom

• SMB: B. The use case must reach into your own machines and historian and no off-the-shelf product fits a typical mixed shop floor; build on cloud or managed tooling together with an integrator, and use a machine builder's built-in parameter assistant where one exists.
• Enterprise: C. Process know-how is core and strategic, the historian data is sensitive, and fleet-wide models across plants justify an internal platform on open-source tooling.
• Developer: C. Ax/BoTorch, Optuna and open OPC UA libraries give full control over models, safety layer and write-back.

Path A — Buy (off-the-shelf)

• Process-industry AI optimisation: Yokogawa (JP, strong EU presence; FKDPP RL-based control), Imubit (closed-loop optimisation for refining/chemicals), AspenTech Aspen DMC / AI-driven optimisation, Honeywell Forge / Profit Optimizer, ABB Ability (CH/SE), Siemens (DE; Industrial AI and Process Insights), AVEVA, Emerson, Schneider Electric (FR).
• Discrete and SMB-friendly: Siemens Insights Hub, TRUMPF (DE) and Engel (AT) smart-process assistants for injection moulding/laser, Kistler (CH) process monitoring, Bosch Rexroth/ctrlX AUTOMATION apps, and MES-embedded optimisation modules. Many machine builders ship parameter assistants with the machine, which is the easiest start for SMBs.
• Choose A when: a vendor has a reference in your process type, you lack data-science staff, and integration into the existing DCS/MES is supported.

Path B — Build on cloud APIs

• Platforms: Azure ML / IoT Operations, AWS SageMaker with IoT SiteWise, Google Vertex AI, Databricks or Snowflake on EU regions; managed time-series stores. Use EU regions and contractual commitments for data residency.
• Optimisation tooling: Google OR-Tools or commercial solvers (Gurobi, FICO Xpress, HiGHS-based services) for constrained search on top of surrogate models; Ax/BoTorch for Bayesian optimisation [5][6].
• Integration: OPC UA / MQTT gateway on-premise, only aggregated data to cloud, write-back stays local through a validated edge component.
• Choose B when: you have a data team and multiple plants, and want fleet-level models.

Path C — Self-host (open source)

• Stack: Ax/BoTorch, Optuna, scikit-learn, XGBoost/LightGBM, PyTorch; RL libraries (Stable-Baselines3, RLlib) with simulators (Modelica/OpenModelica, FMI co-simulation, gymnasium wrappers); time-series stores (TimescaleDB, InfluxDB); open62541 / Eclipse Milo for OPC UA; Node-RED or n8n for flows.
• Hardware: HW1 suffices — tabular and small neural models run on an industrial PC or edge box; GPUs only for RL training or large neural surrogates (HW2).
• Ops effort: needs process engineers plus ML engineers; model validation and cybersecurity (IEC 62443) are the main workload.

Path D — Custom training

Worth it for high-volume or high-value processes where vendor models lack your physics: train plant-specific surrogates, offline RL and hybrid digital twins. Requires months of clean historian data with sufficient excitation (variation in settings); purely observational data often cannot identify cause and effect, so plan deliberate step tests or DoE campaigns.

State of the art (2026)

• Bayesian optimisation and DoE-in-the-loop are mature for ramp-up and recipe tuning; open tools (BoTorch, Ax) are production grade [5][6].
• Offline and safe RL has moved from papers to plant trials: Yokogawa and the chemical producer reported a multi-week autonomous run of a distillation process with the AI directly controlling [1]. Closed-loop AI optimisation products are in commercial use in refining and chemicals [2].
• Hybrid physics-informed models and digital twins reduce data needs and improve extrapolation.
• Foundation-model ideas (pretrained time-series models, LLM copilots for engineers) help with analysis and explanation, not yet with certified control.
• Limits and failure modes: distribution shift (new raw materials, maintenance changes), models exploiting sensor faults, correlation mistaken for causation in historian data, optimising a proxy metric at the expense of unmeasured quality, operator distrust, and cybersecurity exposure of write-back paths. Most plants still run advisory mode; fully autonomous operation is rare and confined to well-instrumented continuous processes.

Shared resources & combinations

• Same stack (tabular-ml + time-series + streaming): PRED-014 soft sensors supply the quality predictions this use case optimises; ANOM-005 anomaly detection feeds the safety watchdog; PRED-006 predictive maintenance shares historian pipelines and edge hardware.
• Planning layer above: OPT-002 scheduling decides what runs when; OPT-009 decides how.
• Vision layer: VIS-001 inspection and VIS-057 in-situ monitoring provide quality feedback.
• Siblings: OPT-031 formulation/blend optimisation, OPT-038 continuous-plant real-time optimisation, OPT-018 natural-language optimisation copilots.

Effort, cost & KPIs

• Path A: weeks to a few months pilot on one line; licence plus integration; vendors often offer pilots with defined value targets.
• Path B: 3–9 months to productive advisory models; cloud cost small versus engineering time.
• Path C/D: 6–18 months, notably for closed loop with safety validation.
• Cost drivers: data cleaning and tag mapping, instrumentation gaps, OT/IT integration, validation and change-control documentation.
• KPIs: first-pass yield, scrap/rework rate, energy and material per unit, throughput/OEE, cycle time, time to qualify new products, share of time AI is active (acceptance), number of operator overrides, specification excursions.

People & change

• Involve early: management (objective, budget), the process owner and process engineers (constraints, validation), operators (daily users and the real experts), IT and IT security (OT/IT segmentation, write-back), data owners of the historian and lab systems. Bring the works council in before the pilot starts.
• Works council: depends. Override statistics, operator IDs and shift logs make the system objectively suited to monitor behaviour, so §87(1) no. 6 BetrVG applies in most plants with a works council; §90(1) no. 3 requires information in good time when work processes are planned with AI. Agree that override data is not used for individual performance rating (staff council in the public sector).
• Data protection: process data is mostly non-personal; if operator data is logged, involve the DPO and check whether a DPIA is needed (Art. 35 GDPR).
• AI literacy: support operators and engineers with short training on what the model can and cannot do (Art. 4 AI Act), see GUIDE-006.
• Pilot and communication: one line, a few motivated operators as champions, advisory mode first. Typical concerns are distrust of black boxes and fear of being replaced; address them with explanations of each suggestion, an always-available off switch and operator veto.
• Human control: engineers approve model releases and limits; changes follow change control.
• Guides: GUIDE-004, GUIDE-008, GUIDE-009.

Risks, regulation & compliance

• EU AI Act: depends. Advisory optimisation is typically minimal risk. An AI system that is a safety component of a product or of critical-infrastructure operation (e.g. management and operation of critical digital infrastructure, or supply of gas, heating, electricity) can be high-risk under Annex I / Annex III; plain efficiency or quality optimisation of production lines is generally not. Assess per plant: if the AI can influence safety functions, treat it as high-risk and apply Art. 9–15 (risk management, data governance, logging, human oversight, robustness) [7]. AI literacy (Art. 4) always applies. No Art. 5 prohibited practice is involved unless operator monitoring is added (emotion inference at the workplace is prohibited).
• Machinery and safety: Machinery Regulation (EU) 2023/1230 (applies from January 2027) addresses machinery with self-evolving behaviour and safety functions; keep AI outside the safety instrumented system (IEC 61511 / ISO 13849 safety functions remain deterministic) [8].
• Cybersecurity: NIS2 for essential/important entities, IEC 62443, Cyber Resilience Act for products with digital elements; segment OT networks and authenticate write-back.
• GDPR: process data is mostly non-personal, but operator IDs, shift logs and override tracking are personal data; involve the works council (Betriebsrat, Germany: co-determination) and avoid using override statistics for individual performance rating.
• Quality regimes: GMP, IATF 16949, ISO 9001 require validated change control for recipe changes.
• Mitigations: hard limits outside the AI, shadow mode before activation, kill switch, drift monitoring, versioned models, full audit trail of every set-point change.

Getting started

1. Pick one process with a measurable quality or energy loss and good instrumentation; write the baseline KPI over 3–6 months.
2. Audit data: tag coverage, sampling rates, lab-result latency, and whether settings varied enough to learn from; plan step tests or a DoE if not.
3. Build an advisory model in shadow mode and compare its suggestions with outcomes for several weeks.
4. Run operator-approved suggestions on one line; log all changes and review monthly with process and quality engineers.
5. Only then consider closed loop: define the safety envelope, fallback and AI Act / machinery risk assessment first.

Sources

1. Yokogawa — AI autonomous control (FKDPP) for chemical plants — https://www.yokogawa.com/library/resources/references/successstory-eneos-materials/
2. Imubit — Closed Loop AI Optimization for process industries — https://imubit.com/
3. AspenTech — AI-driven process optimization — https://www.aspentech.com/en/products/pages/aspen-dmc3
4. Siemens — Industrial AI in manufacturing — https://www.siemens.com/global/en/products/automation/topic-areas/industrial-ai.html
5. BoTorch — Bayesian optimization in PyTorch — https://botorch.org/
6. Ax — Adaptive experimentation platform — https://ax.dev/
7. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
8. Regulation (EU) 2023/1230 (Machinery Regulation) — https://eur-lex.europa.eu/eli/reg/2023/1230/oj