Real-Time AI for Pyroprocess Optimization

Business Overview

Argos, one of the largest cement producers in the Americas, set out to achieve carbon neutrality. To do so, the company defined a roadmap that required profound transformations in operational efficiency, emissions reduction, and the adoption of advanced industrial technologies.

The pyroprocess is the high-temperature stage where raw materials are transformed into clinker inside the kiln, and it is where most of this challenge is concentrated. Producing one ton of clinker requires between 3 and 6 gigajoules of energy, with energy costs accounting for up to 40% of total production costs. A single cement plant can produce up to 5,000 tons of clinker per day, consuming approximately 250,000 liters of fuel daily.

Recognizing the pyroprocess as the critical lever for its decarbonization targets, particularly in a context of increasing carbon pricing and growing pressure to reduce industrial emissions, Argos turned to Barbara to deploy a real-time AI solution capable of optimizing kiln control directly at the edge.

Challenges

When looking at the pyroprocess, the first instinct had traditionally been to improve control using more rules, more tuning, or more supervision. But this approach quickly reached its limits. Argos faced three fundamental challenges:

1. Rule-based systems could not capture the dynamic complexity of the process: with more than 30 variables interacting in a highly non-linear and interdependent way, no set of rules could reliably optimize kiln behaviour.
2. Operations depended heavily on manual interventions: operators continuously adjusted the process based on experience, introducing variability and making performance dependent on individual judgment rather than systematic optimization.
3. Cloud-based approaches were fundamentally incompatible with the process dynamics: critical variations in the flame or kiln load happened in fractions of a second, while cloud latency introduced response times of several seconds, making it unsuitable for closed-loop control in this environment.

Solution

Using Barbara Core, Barbara Panel, and Barbara Marketplace, Argos deployed a three-layer edge architecture that enabled real-time data acquisition, model inference, and closed-loop control, without disrupting existing OT operations. The solution had to not only work reliably under real production conditions, but also scale consistently across every plant in the fleet.

‍

Field Layer (OT Network)

The existing industrial assets, including sensors, PLCs, and control systems operating at Purdue levels 0 to 2, remained fully untouched. A key design principle was not to disrupt existing operations: the architecture connected to them without replacing them, ensuring minimal risk and easier adoption by operations teams. It was hardware-agnostic, able to run on different types of edge devices. At this layer, Kepware OPC UA was configured to read data from PLCs and write it directly to the MQTT Broker on the edge node, acting as the bridge between the OT environment and the edge computing layer.

Edge Layer

This is where the key transformation happened. A distributed computing layer was deployed at Purdue level 3–3.5, managed with Barbara Core and orchestrated remotely via Barbara Panel. Many of the workloads deployed across the fleet were sourced from Barbara Marketplace, where certified industrial applications are available off-the-shelf and ready to deploy at scale. The key components of the system deployed were:

• MQTT Broker: created a unified real-time data backbone, receiving data from Kepware OPC UA and distributing it to all subscribing services simultaneously.
• Scikit Learn Serving: received live process data from the MQTT Broker, ran inference locally in milliseconds, and returned the results back into the broker. Raw data and model outputs were then combined and forwarded downstream for storage, visualization, and alerting.
• ML Monitoring: tracked model performance metrics — including latency and inference counts — using Prometheus, and triggered configurable alerts when thresholds were exceeded, ensuring the model continued to perform reliably in production.
• Alert Manager: received the combined data and model outputs from the MQTT Broker and triggered notifications when anomalies or failures were detected.
• Splunk Ingester: collected all data flowing through the edge and forwarded it to Splunk Cloud via Barbara's built-in VPN service, enabling secure remote access to the corporate analytics environment without exposing the OT network.
• InfluxDB: served as the local time-series database, storing both raw sensor data and model outputs for historical analysis and monitoring.
• Grafana: connected to InfluxDB to provide plant operators with real-time dashboards showing live process data, model predictions, and KPIs, ensuring transparency and building trust in the AI system.

Together, these components formed a fully containerized, microservices-based architecture that enabled modularity and consistency across plants. All services across this layer were deployed and managed through Barbara Panel, which handled deployment, updates, monitoring, and lifecycle management across the entire fleet.

Cloud Layer (Corporate IT)

This layer was used exclusively for non-real-time workloads: historical analysis, model training, reporting, and monitoring. Critically, it was not in the control loop.

Results

By leveraging Barbara's edge AI platform, Argos achieved measurable impact across its Digital Manufacturing Initiative:

1. 38 control loops automated across multiple plants and grinding stations, proving the approach was consistently replicable across different environments.
2. Clinker factor reduced by up to 5%, directly translating into lower CO₂ emissions per ton of cement produced.
3. Specific energy consumption reduced by up to 6%, with a direct effect on both operational cost and environmental footprint.
4. Production increased between 2% and 10%, achieved by stabilizing the process and reducing variability introduced by manual interventions.

Testimonial

"One of the aspects I liked the most about Barbara was its hardware agnosticism, both from the capture side with the wide range of supported industrial protocols, and deployment, covering ARM nodes such as Raspberry Pi and Jetson, as well as Intel CPUs, GPUs, and virtual machines."

— Estefan Wolff, Digital Manufacturing Leader, Argos

Conclusions

By deploying real-time AI directly at the edge of its cement plants, Argos transformed the pyroprocess from a manually controlled, variable operation into a systematically optimized one, achieving simultaneous gains in energy efficiency, CO₂ reduction, and production throughput. What made this possible at scale was Barbara's ability to let industrial teams run, manage, and update AI models across a distributed fleet of plants without requiring them to become IT specialists, keeping operations teams focused on what they do best.

By 2025, Argos had expanded this strategy to more than 90% of its plants, covering key equipment across the entire production chain, bringing them meaningfully closer to their 2050 carbon neutrality objective.

About the Company

Argos is one of the largest cement producers in the Americas. With more than 90 years of history, the Colombian company has established itself as a leading player in the cement industry, becoming the fourth-largest cement producer in Latin America and the third-largest in the United States. Operating in 16 countries with a workforce of more than 4,000 employees, the company continues to expand its global presence while maintaining strong operational performance.

Argos has positioned sustainability as a central strategic commitment, with a defined roadmap toward carbon neutrality. Its cement manufacturing operations are concentrated around energy-intensive processes such as the pyroprocess, where energy costs can represent up to 40% of total production costs, making operational efficiency and emissions reduction key levers in delivering on this objective.

‍