Forge Plant Cognition at Lekki: What Happens When Industrial AI Stops Reporting and Starts Optimizing
Ask a refinery director, a petrochemical plant manager or an LNG terminal lead what is missing today, and the answer is rarely more dashboards. These assets already produce enormous streams of operating data. The binding constraint sits one layer further down: the distance between an alert and a decision an operator trusts enough to act on. That gap is where margin leaks, where an unplanned shutdown quietly takes shape, and where a small drift in operating conditions turns into lost throughput and yield.
It is precisely that gap Honeywell is pointing at with Forge Plant Cognition, an industrial AI platform assembled with NVIDIA AI and Dell Technologies AI infrastructure, aimed at refining, petrochemical and LNG operators trying to create more value from existing assets. Its first public proving ground is Dangote Petroleum Refinery in Lekki, Nigeria, a project described as one of the world's largest single-train refineries.
The design logic: process knowledge inside the model, not beside it
Forge Plant Cognition is layered on the recently launched Forge Cognition platform from Honeywell Technologies. The stated intent is a change of category rather than an upgrade: move past monitoring and responding. Installations in this space typically analyze data and hand back recommendations. The platform instead combines process technology expertise, industrial automation and AI, and fuses real-time operating data with engineering knowledge and the deep process expertise of Honeywell Technologies UOP, grounded in trusted calculations and physics-based models. For a plant team, the practical difference is whether the advice respects what the equipment can physically deliver.
The value proposition is organized around three areas, each of which maps to a recurring management conversation on the shop floor:
- Capture hidden margin. Identify bottlenecks and optimize throughput, yield and product value across interconnected processes, rather than tuning one unit at the expense of the next.
- Protect production. Detect emerging equipment and process issues, recommend corrective actions and help minimize unplanned downtime.
- Scale industrial expertise. Embed specialized process knowledge into AI-powered workflows so operators and engineers make faster, more consistent decisions.
The third area is where more than a century of experience designing and operating refineries inside Honeywell Technologies UOP does the heavy lifting. The premise is unglamorous but accurate: even small changes in operating conditions affect efficiency and throughput. Specialized AI capabilities draw on integrated intelligence built from that knowledge base to identify operational constraints and recommend actions to improve plant performance.
Case in focus: what the Dangote deployment actually offers an operations team
The Lekki Refinery engagement is worth reading as a template rather than as announcement noise, because it pairs a performance thesis, a capability thesis and an unusually explicit caveat.
Performance. Dangote Petroleum Refinery will integrate operating data with process expertise from Honeywell Technologies and AI-powered optimization to identify constraints, accelerate troubleshooting and enable more informed decisions. Based on preliminary data, Honeywell Technologies estimates that AI-enabled optimization with Forge Plant Cognition could deliver operational benefit of more than 29 cents a barrel at the facility, and that at full-capacity throughput this could generate more than $60 million in gross benefit per year.
Capability. Beyond output, the platform supports Dangote's commitment to develop a highly skilled local workforce. By embedding specialized process knowledge into AI-powered tools, engineers and operators gain access to expert guidance, faster problem-solving and build technical capabilities. As hiring continues, these tools underpin the goal of a workforce that is 70% Nigerian and a steady reduction in reliance on external expertise.
Scope. Building on nearly a decade of collaboration, the two companies say Honeywell Technologies is helping the refinery increase capacity from 650,000 to 1.4 million barrels per day within three years by optimizing existing assets and speeding up market delivery.
Caveat. Honeywell Technologies states that the estimate of more than $60 million and more than 29 cents a barrel is not guaranteed, is specific to this facility, and is gross of implementation costs. Actual results depend on refinery configuration, operating conditions, crude slate, utilization, baseline performance, data quality, product margins and customer implementation.
Reading the business case: which assumptions transfer
| Assumption to test | What the vendor discloses | Why it changes the arithmetic |
|---|---|---|
| Facility specificity | The estimate is specific to this facility | A per-barrel figure modelled on one train, one configuration and one crude slate is not portable to another asset without a new baseline. |
| Cost basis | The figure is gross of implementation costs | A net business case requires the deployment, integration and services spend alongside the benefit. |
| Operating variables | Configuration, operating conditions, crude slate, utilization, baseline performance, data quality, product margins, customer implementation | Each of these can move throughput and yield optimization; weak data quality or an unrepresentative baseline erodes the claim first. |
| Guarantee status | Explicitly not guaranteed, preliminary data | It is an indicative estimate, so procurement and finance teams should treat it as a hypothesis to validate in a pilot. |
Technology stack, layer by layer
The platform is powered by Dell AI Factory with NVIDIA and supported by Dell Services. The individual layers matter to engineering and IT leaders because each one answers a different operational question.
| Stack layer | Named component | Function in the plant workflow |
|---|---|---|
| Data-understanding agents | NVIDIA Nemotron open models | Understand and interpret operational data. |
| Simulation environment | NVIDIA Omniverse libraries | Create a digital twin of the plant and its processes for simulation-based training and planning. |
| Vision AI | NVIDIA Cosmos open models with the NVIDIA Metropolis VSS Blueprint | Enable vision AI agents for real-time alerts and detection of anomalies. |
| Physics AI | NVIDIA PhysicsNemo framework | Build, train and fine-tune efficient and performant physics AI models at scale to support digital twin integrated agile plant optimization workflows. |
| Accelerated compute | NVIDIA RTX Pro 6000 and NVIDIA RTX Pro 4000 | Run these applications; NVIDIA supplies the accelerated computing capabilities behind the platform. |
Common misconceptions to clear before a pilot
- Industrial AI is a reporting layer. Dashboards that stop at recommendations leave the action gap untouched. The stated ambition of platforms like Forge Plant Cognition is to translate insight into operational outcomes.
- Process knowledge is optional if the model is large enough. The differentiation described here rests on trusted calculations, physics-based models and Honeywell Technologies UOP expertise, not on raw data volume.
- A published benefit figure is a contract. The Lekki estimate is preliminary, facility-specific, gross of implementation costs and explicitly not guaranteed.
- Unit-level optimization equals plant-level gain. The margin case is built on interconnected processes, where a bottleneck moved rather than removed still costs throughput and yield.
- AI substitutes for scarce expertise. The deployment narrative is closer to distribution than substitution: embedding specialized process knowledge so more engineers and operators can reach expert guidance and build capability locally.
Action recommendations for managers evaluating this category
- Fix the baseline first. Document current configuration, operating conditions, crude slate, utilization, baseline performance and data quality before any per-barrel claim is evaluated.
- Model net, not gross. Put implementation and services costs beside the benefit estimate so the case survives a finance review.
- Map constraints across interconnected processes. Target the hidden margin that appears when throughput, yield and product value are optimized together, and pair it with the unplanned downtime use case so protection and capture are measured in the same pilot.
- Require physics-grounded outputs. Insist that recommendations respect trusted calculations and physics-based models, otherwise operators will not act on them.
- Treat the platform as a training channel. Use embedded process knowledge to grow local engineering and operations capability, which is the durability argument in the Lekki case.
- Keep a human in the loop on corrective actions. Emerging equipment and process issues still need accountable decisions inside the operating procedure.
FAQ
What is Forge Plant Cognition?
An industrial AI platform from Honeywell built with NVIDIA AI and Dell Technologies AI infrastructure, layered on the Forge Cognition platform, and aimed at refining, petrochemical and LNG operators.
Where is it deployed first?
Dangote Petroleum Refinery in Lekki, Nigeria, described as one of the world's largest single-train refineries, with the stated project aim of lifting throughput while strengthening local engineering and operations teams.
Which parts of the stack do the heavy work?
Dell AI Factory with NVIDIA and Dell Services host the workloads; NVIDIA Nemotron handles operational-data understanding, NVIDIA Omniverse builds the digital twin for simulation-based training and planning, NVIDIA Cosmos with the NVIDIA Metropolis VSS Blueprint supports vision AI alerts, NVIDIA PhysicsNemo covers physics AI models, and the applications run on NVIDIA RTX Pro 6000 and NVIDIA RTX Pro 4000.
Should a benefit of more than $60 million per year be treated as a planning number?
No. Honeywell Technologies ties more than 29 cents a barrel and more than $60 million to preliminary data at this specific facility, gross of implementation costs, and not guaranteed. It is a hypothesis for a scoped pilot, not a commitment.
Why does the workforce angle matter to operators?
Because scaling expertise is one of the three value areas: embedding specialized process knowledge into AI-powered workflows is expected to make decisions faster and more consistent, while supporting a workforce that is 70% Nigerian and reducing reliance on external expertise over time.
For managers who have already invested in sensors, historians and analytics, the interesting signal from the Lekki Refinery is not the headline number. It is the shift in placement: moving AI out of the reporting layer and into the constraint-and-action layer, where throughput, downtime and local capability are actually decided. The capacity build-out now targeted at 650,000 to 1.4 million barrels per day within three years will show whether a per-barrel thesis built on physics-grounded models holds beyond the preliminary data that produced it.
Editorial note — Industrial Weekly / WorldIndustrialWeekly. Byline: Michael Torres. Contact: news@industrialweekly.com.