ProductionSpecialized caseVerified Real deployment at a named or precisely described company, with a measured result from a primary source.

AI adjusts the cement kiln in real time to burn less fuel and keep clinker quality steady (Heidelberg Materials, SECIL Sibline, Holcim)

Added 2026-10-23 · Last updated 2026-10-26

Manufacturers in this case: Heidelberg Materials · Holcim · SECIL · all manufacturers using AI

Role

Production (continuous process control: kiln and pyroprocess optimization) · Typical users: process engineer, kiln or control-room operator, production manager, plant manager, quality and laboratory lead, energy or sustainability manager

The problem

A cement kiln is a hard process to run steadily. Raw meal chemistry varies, and plants burn more and more alternative fuels such as refuse-derived fuel (fuel made from sorted waste), whose calorific value (energy content) changes from one load to the next. Operators and advanced process control (APC) keep correcting fuel, air and feed, but each shift reacts after the fact: the kiln is run with safety margins that cost fuel, clinker (the intermediate product that is ground into cement) quality swings, and process engineers spend their time on manual tuning and shift-to-shift troubleshooting. Heidelberg Materials' Mokra plant (Czechia) already ran a thermal substitution rate (the share of kiln fuel energy that comes from alternative fuels) of up to 86%, mainly with refuse-derived fuels, on top of ABB Ability Expert Optimizer, and wanted to extend that APC with AI.

How AI helps

Machine-learning models trained on the plant's live process, laboratory and chemical data act as soft sensors (software estimates of values that are slow or hard to measure directly): they continuously predict key values (for example clinker quality indicators such as C3S, the main strength-giving mineral in clinker, or free lime, unreacted lime left in it) and recommend or set the next control moves for fuel mix, feed, kiln speed and air. At Mokra the AI sits on top of the existing ABB APC and gives it dynamic targets in closed loop (its moves are applied directly, without waiting for an operator); at Sibline the optimizer runs the kiln in continuous autopilot; at Holcim's Siggenthal plant an AI digital twin (a model that mirrors the kiln's live behavior) balances fluctuating alternative-fuel calorific values and stabilizes the flame in autopilot. Operators keep the constraints and can take back control at any time. Note that the set-ups differ: Mokra's AI sits on top of an existing APC, while the Gigaton system announced in May 2026 (the new name of Carbon Re) also controls the actuators directly.

Step by step

Measure the baseline over several weeks: specific heat consumption, fuel cost per tonne, thermal substitution rate, throughput, and the variation of key quality values such as free lime and C3S, shift by shift.
Check the data: historian tags, lab results with timestamps, fuel deliveries and their calorific values; soft sensors are only as good as the lab data they learn from.
Start in advisory mode: the model predicts quality and recommends moves, and operators compare them with what they would have done.
Agree the operating limits with operations and safety (temperatures, oxygen, CO, pressure), and keep the existing APC and interlocks in charge of them.
Run a fair test, as Mokra did: alternate periods with and without the AI and correct for fuel mix and raw material quality before claiming a saving.
Move to closed loop step by step, with a clear way for operators to take back control, and review the results monthly with the plant team.

Tools

Carbon Re, now Gigaton (AI process control on top of ABB Ability Expert Optimizer or FLSmidth PXP): used at Heidelberg Materials Mokra
ES Processing Kiln Process Optimizer (AI and APC): used at Sibline Cement, SECIL Group
Aix Process KILN AIxperT (AI digital twin of the kiln): used at Holcim Siggenthal
Other tools to evaluate: the AI or APC modules of your control-system vendor (ABB, FLSmidth, Siemens), and process analytics tools that read your historian and lab data
For a first offline study: any LLM with a data-analysis or code tool (ChatGPT, Claude, Copilot, Gemini) on an export of past shifts, with the prompt below

In the AI tools index: Aix Process KILN AIxperT · Carbon Re (now Gigaton) · ES Processing Kiln Process Optimizer · all AI tools used by manufacturers

Result

Heidelberg Materials, Mokra plant (Czechia), Carbon Re AI (now named Gigaton) integrated with ABB Ability Expert Optimizer: in the first month of continuous operation, fuel cost index down 4.1%, fuel-derived CO2 down 4.5 kg per tonne of clinker (2%), and C3S variance down 33%; in a one-month on, one-month off test controlling for fuel mix and raw material quality, specific heat consumption down 2.2%; in a Carbon Re-sponsored CemNet article (2 June 2025) the plant director says the live connection and dashboard were delivered in 8 weeks (that is the connection and dashboard, not a full roll-out or the time to the results above) (University of Cambridge Enterprise news and the sponsored article). Sibline Cement (SECIL Group, Lebanon), ES Processing Kiln Process Optimizer: kiln throughput up 2%, thermal energy consumption down about 1.8%, specific electrical energy down 2.8%, free lime standard deviation down 48.5% and C3S variation down 5% (the vendor page gives no period or comparison base), with the kiln in continuous autonomous operation (vendor news, featured in World Cement, October 2025). Holcim (Schweiz), Siggenthal plant, Aix Process KILN AIxperT: deployed to balance alternative-fuel calorific values, predict clinker quality in real time and stabilize the flame in autopilot at high substitution rates; no figures published, and no measured result or duration of operation in the source (CemNet news, 6 March 2025).

As reported by the source; most figures are vendor or self-reported and not independently verified.

Human review: cautions

Operating limits, interlocks and safety stay with the existing control system and the operators; AI recommendations and closed-loop moves work inside those limits, and operators can take back control.
Measure savings fairly: fuel mix, raw materials and production rate change all the time, which is why Mokra used a one-month on, one-month off test.
Soft sensors drift when raw materials or fuels change; check them against lab results continuously.
The Mokra figures come from the university spinout's news and vendor material (one of the pages is a sponsored article) with a quote from the plant, so the five sources are not five independent confirmations; the Sibline figures are published by the vendor; Holcim published no figures. Results from one plant do not transfer directly to another kiln.
This case is for cement plants that already have a kiln APC, a historian and reliable lab data; it is a specialist case, and costs, integration effort and net savings in money are not published.

Try it yourself

You are an experienced cement process engineer.

I will paste an export of past kiln operation, one row per hour or per shift: [timestamp, kiln feed rate, fuel rates by fuel type, alternative-fuel calorific value if known, kiln speed, main drive current, back-end and burning-zone temperatures, O2, CO, NOx, secondary air temperature, lab free lime, lab C3S, clinker production], plus our operating limits.

Do this:
1. Describe the data: which columns are missing or have gaps, how often lab results arrive and how they line up with process data.
2. Find the periods with high free lime or high fuel use per tonne, and list what was different in the hours before (fuel mix, calorific value, feed, O2, kiln speed), quoting the rows.
3. Compare the best and worst 10% of shifts by fuel per tonne, with the average of each variable, and say which differences look like causes and which may be effects.
4. Suggest 3 to 5 hypotheses for steadier operation that we could test in advisory mode, each with the variable to watch and the limit that must not be crossed.
5. List the data we would need before trusting any model in closed loop.

Write any calculation as a spreadsheet formula or code I can run myself, and mark every hypothesis as unproven. Never suggest moving outside our operating limits.
Use only the data I paste in; if something is missing, write 'not found' or list it as a question instead of guessing. Do not send, order, post or change anything in any system: produce drafts for my review and approval.

Want more like this? Join the newsletter

Related cases

All 12 Production use cases →