Severstal: 'Adelina' AI Operator Boosted Rolling Mill Throughput by 5% with Multi-Billion Ruble Impact
The 'Adelina' AI controller increased Hot Strip Mill 2000 throughput by roughly 5% with zero capital equipment expansion—unlocking over 300,000 additional tonnes of rolled steel per year.
| Company & Sector | Северсталь (Severstal) — Metallurgy & Heavy Industry |
|---|---|
| Models & AI Stack | Closed-Loop Industrial ML · PLC / SCADA Integration · CNN Surface Defect Vision · Time-Series Telemetry · Gradient Boosting · Predictive Maintenance |
| Rollout Timeline | 2019–2020 — launch of 'Adelina' closed-loop AI on CherMK Mill 2000 and Kaggle Steel Defect dataset release; 2021–2023 — expansion of ML autopilots to pickling and galvanizing lines; 2024–2026 — GenAI assistants for maintenance and metallurgy engineers. |
| Verification | Verified 2026-10 · 3 primary sources |
Business Context & Scale
Severstal is one of Russia's largest vertically integrated steel and mining companies, anchored by the Cherepovets Steel Mill (CherMK) producing over 11 million tonnes of steel annually.
In integrated steelmaking, hot strip rolling mills (Mill 2000) often act as the primary throughput bottleneck. Every second of inter-slab gap or suboptimal rolling speed directly caps annual finished steel output.
Severstal Digital became one of the first industrial AI teams globally to move beyond advisory dashboards to closed-loop real-time AI machine control on heavy metallurgical equipment.
Business Problem & Bottlenecks
Historically, human pulpit operators manually controlled slab discharge pacing and rolling speeds. Because cobbles (strip jams inside rolling stands) cause expensive downtime, operators naturally added conservative safety buffers between slabs.
Yet every slab differs in steel grade, furnace temperature, chemistry, and target gauge—making it impossible for a human to solve multi-variable load and thermal equations every millisecond.
A second challenge is surface defect inspection (scale, cracks, slivers) on steel strip moving at 15–20 meters per second, far beyond human visual perception.
AI Architecture & Implementation
Severstal Digital engineered and deployed 'Adelina'—a closed-loop ML control system on Hot Strip Mill 2000. Reading hundreds of live sensor streams from reheating furnaces and rolling stands, the model directly governs slab discharge timing and roller table speeds without manual operator intervention.
The algorithm computes the minimum safe inter-slab gap tailored to each slab's exact steel grade and thermal profile, preventing collisions and motor overloads.
Simultaneously, deep convolutional Computer Vision pipelines inspect moving steel strip at full line speed, classifying over 20 surface defect types—technology later shared with the global ML community via the Severstal Steel Defect Detection benchmark on Kaggle.
“"As long as a model merely flashes a hint on a screen, operators ignore it half the time out of habit. The real industrial breakthrough happens when the algorithm enters closed-loop control and drives the machinery directly." — Severstal Digital
Measurable Impact & Economics
The 'Adelina' AI controller increased Hot Strip Mill 2000 throughput by roughly 5% with zero capital equipment expansion—unlocking over 300,000 additional tonnes of rolled steel per year.
Across its portfolio of industrial AI deployments (rolling optimization, blast-furnace burden optimization, predictive maintenance, and surface vision), Severstal achieved multi-billion-ruble annual EBITDA impact.
Real-time surface defect vision significantly reduced defective coil shipments to automotive and construction customers while catching process drift early.
Key Takeaways for Industry
In heavy industry, closed-loop PLC actuation delivers 3x–5x the financial impact of passive operator-advisory screens.
Unlocking 5% more throughput at a plant's primary bottleneck via software generates returns comparable to major CapEx expansion.
Hardware safety interlocks inside the PLC guarantee fail-safe operation when ML predictions approach physical limits.
Open-sourcing an anonymized defect dataset on Kaggle crowdsourced state-of-the-art segmentation architectures from thousands of global researchers.
Verified Primary Sources
Key Questions Answered
What happens if the AI model issues an unsafe speed command during rolling?
Deterministic safety interlocks inside the industrial PLC sit between the ML server and the drives, rejecting any command that violates thermal or current bounds.
Why did a steelmaker publish its defect imagery on Kaggle?
Over 2,400 global teams competed, yielding open-source segmentation architectures that accelerated Severstal's production defect-detection accuracy and inference speed.
Did deploying 'Adelina' require replacing physical rolling mill machinery?
No. The system connected to existing SCADA/PLC sensors and drives on Mill 2000, achieving the 5% throughput gain purely through software optimization.
Delegate Deep Research, Coding & Multi-Step Workflows to Nomad Agent
Deploy autonomous task execution with Nomad Agent (agent.nomadnet.ai) — automated web research, data synthesis, code generation, and MCP tool orchestration.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.