1st Global CementAI Conference & Exhibition, 19 - 20 May 2026, Brussels, Belgium

The first Global CementAI Conference & Exhibition has successfully taken place in Brussels, with approximately 95 delegates from over 20 countries, 20 presentations and 12 hours of networking. The 2nd Global CementAI Conference will take place in Munich in May 2027.

Day One

The conference started with event organiser David Perilli, Global Cement Magazine, giving a short history of artificial intelligence (AI) and robotics in the cement industry. He ran through the various forms of AI and its methods - including model predictive control, neural networks, the internet of things, the use of predictive maintenance, digital twins, deep learning and systems learning for themselves – and how they link to industrial control hierarchy. He pointed out a number of problems with AI systems, including the garbage-in, garbage-out problem, data scarcity, the black box problem, hallucinations, retraining models, costs and cultural and operational differences between programmers and chemical engineers. Perilli also covered the use of robotics in the sector, including the use of driverless trucks, drones and walk-about robots. AI can also be used to optimise logistics, generating new cement formulation, human-machine teaming, computer vision, knowledge management and customer service and back-office functions. Safety frameworks need to be codified, cybersecurity needs to be tightened and the legal frameworks for the use of AI need to be clarified so that it is clear where any liabilities lie.

Kevin Treiber, VDZ introduced the association’s strategy to develop a new data model for the cement sector. This approach uses standardised virtual representations of assets, measurement points and meters, integrates with existing systems and represents organisational information using graph theory. It is intended to improve data accessibility by creating a ‘single point of truth,’ which then allows different cement plant staff to view information in a way that suits their job roles. The longer term aim is to create an open and standardised data model that enables a plug-and-play architecture for AI tools, making it easier to compare plants and to deploy applications to new plant layouts.

Christian Frentz, INFORM explained how his company uses a wide variety of AI and software approaches to tackle logistics issues for cement and concrete companies. Its products help customers optimise fleet scheduling in real time, ultimately increasing truck productivity. Frentz went on to reveal how INFORM can also help to use dynamic pricing, not to maximise profits directly, but instead as a cost efficiency tool to smooth a customer’s order book and use resources better.

Mélissa Malischewski, SiloConnect detailed how her company’s sensor-based product measures the mechanical deformation of the legs of silos. This information is then processed to train one AI model per silo and provide insight into logistics and then sales on how they are used. In the example given for Holcim Spain, for example, last minute orders and transport lead time were reduced, and mean truck delivery volumes were increased.

Marius Sugentaitis, OreScanner presented his vision-based AI system that uses visual and infra-red (IR) spectrum camera data combined with operator decision data, by recording control panel data, to assess the quality of ore being fed into a crusher at a lime plant. The system now controls the conveyor in a closed-loop feeding the crusher continuously, deciding whether to accept ore for crushing or stockpile it. A lime plant in Germany has been running the new tool since mid-2025. An additional 3% or 15,000t/yr of material is now being accepted for crushing, saving the plant money.

Andres Duque, Ripik.AI talked about his company’s visual based AI system to estimate the calorific value of alternative fuels (AF). It uses visual and IR cameras to check for material composition, particle size and moisture content. In one example, Duque said the proportion of unburnt AF decreased from 6% to after 3% after adopting the Ripik.AI system. During audience questions, Duque added that AF samples had been taken during the training phase to work out the various calorific values.

Mareike Hoyer, ci-tec next spoke about the possibility to quantify free lime during clinker production using cooled kiln-head infrared cameras which can capture temperatures from the kiln even through dust. Each pixel can be used as a temperature measurement, and through human and automatic analysis, the free lime can be deduced. However, in a single moment, high free lime and low free lime can look identical. Mareike suggested that we need to “watch the whole movie” to be able to understand the situation, which ci-tec does through analysing a sequence of infrared images using a temporal deep learning model using convolutional recurrent neural networks (taking into account past history) to deliver a final value. The information from the recordings is matched with a ‘ground-truth’ measurement from the laboratory. “Soft sensor prediction catches free lime trends early,” said Hoyer. The system manages to predict free lime accurately, with only 3% misclassification in low- and target-quality material. The system improves product consistency and quality.

Javier García, Optimitive next spoke about the use of AI agents in cement plant optimisation. A non-linear approach was taken to real-time optimisation (RTO), using machine learning algorithms in order to optimise economic key performance indicators (KPIs). Javier spoke about the creation of autonomous AI agents, that can be set to achieve many tasks.

Kenny Wong, Carbon Re started by admitting that earlier attempts to place AI systems on top of existing advanced process control (APC) systems at cement plants have often failed, as these systems are rarely maintained well enough to cope. Carbon Re’s approach has been to build an AI system that uses self-learning control. It starts by gathering data, builds a digital twin, creates controllers and then takes action. It executes control in a wide range of states, maintaining performance over time autonomously. It gives control of the actuators to the AI and updates control from feedback. Crucially, the company also gives feedback to the models by labelling data, in coordination with human operators, to help explain subsequent actions the AI recommends and avoid black box issues. In this way, a library of label data is collected and a dashboard created that can offer insight into how and why a particular AI model decision has been made.

Mathis Ricker, KHD spoke about his company’s simulation and process optimisation tools. The Simulex operator product provides full plant simulation, based on mechanistic models and continuous refinement. It is a hybrid AI approach using mechanistic or physics-based models to start with, that are then adjusted by a data driven model. Using the example of the power consumption of roller press, Ricker demonstrated that the mechanistic part suited dynamic variation better, whilst the data part suited the trend better. The combined output matched the real-world data much better than either part alone and reduced the modelling error from 3.47% to 0.38%. Similarly, KHD’s process optimiser uses a hybrid AI determining setpoints in a closed-loop system and uses an open-loop system to recommend actions to human operators. A two week performance test at a cement plant in China reportedly reduced specific thermal energy consumption by 2%, improved the quality of clinker free lime and saved the plant €0.5m/yr by recommending 15 optimised setpoints.

Anders Noe Dam, Fuller Technologies ended the presentations for the first day by introducing his company’s APC product ECS/ProcessExpert (PXP). The latest version PXP 9.1 deploys self-adaptive controllers to deliver a reported improvement range of 4 – 7% for kiln or mill specific heat consumption, production and specific power consumption from manual PID control. The system has had over 150 sales since its launch in mid-2024, with at least 50 installations. However, Fuller has now started a partnership with Imubit to add AI softsensors to start making real-time predictions in a closed loop, allowing for proactive control. This has started by estimating gas levels in the kiln inlet, clinker free lime and Blaine fineness at the mill at the Grupo Cementos Portland Valderrivas (GCPV) Els Monjos plant in Spain. A 17% reduction in the standard deviation of free lime was reported, among other improvements. The two companies anticipate a 5 – 10% improvement using PXP and AI softsensors in a closed loop.

Global CementAI Awards

At the end of the first day, the conference delegates gathered for the Global CementAI Awards. Holcim won the award for AI-user of the year. CemAI won the award for AI supplier of the year. Fuller Technologies won the award for automation supplier of the year. Holcim and C3 won the award for AI project of the year for their work on a predictive maintenance project. Heidelberg Materials and Pronto won the award for automation project of the year for their work on an autonomous haulage project. The inaugural Global CementAI Personality of the Year was awarded to Dirk Schlemper, INFORM.

Day Two

The second day of the conference started with a panel discussion on AI adoption in the cement sector, including experts from a cement producer, an equipment manufacturer and an AI company.

Matthias Schumacher, aixprocess gave the first paper of the day with a study co-authored by Michael Suter, Holcim exploring how an AI-based kiln control system had been used at the Siggenthal cement plant. Schumacher emphasised that malfunctions on a pyroprocessing line resulting from AF usage should be solved before AI models are applied. Aixprocess’ hybrid AI approach uses a digital twin that then uses an AI model for process settings and a process housekeeping approach to determine technical implementations. It also uses SHAP (Shapley Additive explanation) values to explain where the reasons for its predictions come from, such as a free lime estimate. Tests at the cement plant using the Kiln AIxperT system used the system with recommendations either manually implemented or in a closed loop control mode, in late 2024 to mid-2025. After further training and optimisation, the system delivered a reduction in specific thermal energy consumption of 150 – 200kJ/kg of clinker by late 2025.

Shervin Sabzevari, KIMA Process pointed out that cement production involves nonlinear thermochemical processes, has time-varying operating regimes and has safety-critical actuation. This is why, in his view, AI deployments have been predominantly hybrid so far. Closed-loop control required bounded and interpretable architecture. He noted that neural network soft sensors are good at predictive estimates and correlation-based modelling, but are depending on stable training data and have limited interpretability. Subsequently, using historical data can be ‘fatal’ if purely data-driven models are given closed-loop control due to operating changes, accumulated drift and prediction bias. Sabzevari recommended that machine learning predictions inform, rather than control processes. He was also keen to note that the AI label has become very broad, and that vision-based systems should not necessarily be evaluated with the same scrutiny as those that run processes.

Reiner Gnauert, ScrapeTec explained how the E-PrimeTracker product uses predictive maintenance to reduce maintenance for operators using (non-AI based) mathematical models. It gathers data from a distance sensor, measures the Hall-Effect and uses load cells to process horizontal tilt, vertical tilt, belt speed and applied forces variables. It then builds a distance map, creates segments or bins, builds a ‘healthy’ operational model and then uses all of this to compare live data against the model to find anomalies and attribute them to a cause, for example, such as a badly aligned belt splice. Alerts can then be sent to operators.

Ioannis Sandis and Tejas Maru, CemAI took it in turns to first describe their company’s process optimisation and predictive maintenance products. The organisation is an affiliate of Titan and has predictive maintenance products installed in all of the cement producer’s 12 plants, with more references being commissioned in the first half of 2026. Its process optimiser product uses AI to look for new combinations of setpoints to increase efficiency. In a case study looking at process optimisation, throughput was increased by optimising preheater O2 levels and 5^th stage cyclone temperate. This in turn led to a reduced free lime standard deviation and reduced specific heat consumption in the main burner by 3%. Installation of the product takes eight to 12 weeks. Case studies demonstrating early issues found at a kiln pier and a finish mill leading to the avoidance of downtime and better Blaine targets were detailed.

Raghuraj Rao, AKXA Tech pointed out that cement plants worldwide are ageing. He noted that optimising one piece of equipment, such as the cooler or the mill, will upset other parts of the system. The challenge for AI is to optimise the whole system. “Fight with the known devil, before loving the unknown God,” he said, and “use simple OPEX interventions before going for big CAPEX.” He insisted that process engineers can only effectively increase the efficiency of the whole set-up if they decrease variability first. Variability comes from multiple sources: control-related issues; cascade loops; valve/damper/feeder operations; design limitations; external disturbances; and transmitter/sensor issues, which can all cause problems. “Before you try to optimise, you must stabilise,” said Rao. Identifying the cause of variability, through a ‘fluctuation audit’ and addressing it, will deliver up to 60% of the benefit of applying a much higher level AI to the process. Variability is natural, but a flat heartbeat is not healthy, he said. Rao was the first speaker at the conference to mention natural intelligence (NI): the use of human wisdom. He prefers the augmentation of NI with AI.

Sean Maley, Infinite Uptime described how his company goes beyond predictive maintenance, and delivers prescriptive maintenance, which includes the detection of possible future problems, and then provides actionable advice. Feedback from the plant is fed back into the system, so that the system’s ability to detect and predict can improve over time. The system is based on an analysis of all possible failure modes using existing plant data, which is screened by a reliability intelligence engineering team to determine what is likely and important, and - knowing the sensor signatures of any possible failure mode and cause - the resultant data is used to help with root cause analysis and automated report production. Robert Christinger of Switzerland’s Vigier Cement stepped up to report, first-hand, on how the system had saved his cement plant hundreds of thousands of Euros in avoided downtime.

Then Haytham Sobhi, ASEC Automation spoke about the current advantages and disadvantages of AI in the cement industry. Haytham pointed out that there is a 28-day blind spot in the industry, since compression tests take this long to complete. The blind spot leads to an expensive over-specification in the cement. AI may have a role in predicting 28-day compressive strength using soft sensors in a virtual laboratory, eliminating over-design. Colleagues retiring and taking their knowledge is a continuing problem potentially addressable by AI, while undetected anomalies can escalate into ‘catastrophes,’ to which AI may struggle to respond. Haytham pointed to the ‘trust gap’ of black-box unexplainable AI, and also the risk of cognitive atrophy, through 'over-reliance' on ‘autopilot.’ He reiterated that keeping AI accurate - a requirement due to plant and process changes - needs expertise and resources, and that it may become expensive over time. “Implementing AI on an unstable system is like putting a golden roof on a shack,” he said.

Jessica Burley, Juna.AI introduced how her company is using AI agents to process cement plant data and make it accessible to staff using natural language. The software can capture data from a wide variety of standardised plant sources, document types and cope with less structured raw data. Its Agentic Factory OS product then connects and then builds a semantic layer to understand and act on the data. It is intended to allow engineers to act like data scientists and run queries on the data without necessarily being forced to use formal database syntax such as SQL. The company is currently working with an unnamed cement company, where it has automated 150 processes across six functions. This has allowed process engineers, quality engineers and reliability engineers to each scrutinise the data and build custom reports.

Finally, Alexandre Ouzia, Heidelberg Materials presented his research on quantifying the phase assemblage of low carbon binders using machine learning and scanning electron microscopy (SEM). The goal is to create a workflow to identify the characteristics of new SCMs faster than existing approaches. Developed as part of the EU-supported MatCHMaker project, it uses an image clustering approach to segment, isolate and characterise individual mineral phases from SEM-EDX pixel data matrices. The system automatically measures diameter, surface area, roughness, aspect ratio and particle size distribution. As the hydrates are mixed, it can be challenging to visually separate them. However, this project uses linear spectral unmixing to determine the ratio of each mineral phase. Further work is required on refining the clustering algorithm.

Prizes and farewells

At the end of the conference programme, delegates met once more to say their final goodbyes. After delegate voting, the best presentation awards were announced: Raghuraj Rao was third for his paper on decreasing variability in process optimisation; Jessica Burley was second for her paper on capturing expert knowledge and making it accessible; and Alexandre Ouzia was first for his presentation on using machine learning to identify mineral phases.

The event was highly praised by delegates for its papers, networking and time-keeping. The 2^nd Global CementAI Conference will take place in Munich, Germany in May 2027.