Confidential

Chapter 05 · Main Report

Project Roadmap¶

Portfolio Overview¶

The Innovation Sprint identified 175 initiatives across four sites, consolidated into 61 site-level projects and 11 cross-site programs. After scoring each against six evidence-based criteria (cross-site scalability, quick-ROI entry point, executive endorsement, business case strength, prove-then-scale path, and scope discipline), 10 projects qualify for the roadmap. One additional project (PRJ-02: Production Scheduling) is deferred to program expansion.

Rank	Project	Horizon	Score	Sites	Aggregate Value
1	PRJ-09: Knowledge Capture / Virtual SME	H1	4.30	5	$1-5M/site + risk mitigation
2	PRJ-07: Intra-Plant Logistics	H1→H2	4.25	4	$22-60M
3	PRJ-03: PdM Platform	H1→H2	3.70	4	$3-24M/site
3	PRJ-04: Quality & Yield	H2	3.70	3 (steel)	$15-43M
5	PRJ-05: Cobble & Process Risk	H2	3.55	3 (steel) + 6 BFs	$8-35M
6	PRJ-01: Ops-Maint Data Integration	H1	3.30	5	$2-5M/site
6	PRJ-10: Process Chemistry Optimization	H1→H2	3.30	1 (mine) + analog	$8-18M/yr
8	PRJ-06: Maintenance Workflow	H1	3.10	4	$2-19M/site
9	PRJ-08: Caster Chemistry & Steelmaking	H2	3.05	3 (steel)	$11-33M
10	PRJ-11: Coke Plant Ops & Battery Vision	H1→H2	2.70	1 + 3 coke plants	$7-17M

The projects are organized below by five thematic areas. Within each area, projects are presented in horizon sequence.

Operational Intelligence¶

Closing the foundational information loops that every other project depends on.

PRJ-01: Ops-Maintenance Data Integration¶

The problem. Operations and maintenance at every CLF site track the same events in different systems using different naming conventions. A 14-minute production delay logged by operations may have no matching work order in the CMMS. Delay categories are corrected after the fact. Monthly reliability metrics exist but are never connected to the maintenance actions that drive them.

The result: maintenance runs reactive because there is no systematic feedback loop between what breaks, why it broke, and what was done about it.

Validated at: Cleveland (strongest signal — every transcript, every stakeholder), Middletown (Steve Longbottom + John Houston championed at Day 5 readout), Tilden (Pete Austin + George Harmon), Burns Harbor (Miles B + BF Process Engineer), Indiana Harbor (worst-case validation — 282 unread PdM emails, 4 radio channels, no shared information).

Phase 1 (H1, Months 1-4): Entity extraction and rule matching between ops delay categories and CMMS work order types at entry site.

LLM (via Azure AI Foundry) extracts equipment IDs, failure modes, and timestamps from free-text delay logs and work order descriptions. Deterministic matching on extracted entities links the two systems. Tabware's equipment hierarchy provides validation (extracted entities checked against known equipment).
At Cleveland: direct SQL extraction from the underlying delay report database (bypassing web dashboards) paired with Tabware CMMS data.
Misattribution analysis combines descriptive analytics (cross-referencing timestamps, categories, equipment to surface systematic gaps) with unsupervised clustering to find misattribution patterns that manual analysis would miss.
Unified daily dashboard with role-based views: maintenance planners see work order gaps (which delays have no matching maintenance activity), operations leadership sees corrected delay attribution for shift/department performance metrics. Same data model, different lenses.

Phase 2 (H1-H2, Months 5-8): Extend to second site with automated delay categorization.

At Middletown: flat-file export from the Armco-era IBM Mainframe (Teams/SWAMI) — all integrations through an abstraction layer so they survive any future platform migration.
At Tilden: ETL through the data platform (Databricks/Fabric) connecting Modular dispatch data to ELLIPS.
Automated delay categorization: a supervised classification model trained on the cleaned and validated Phase 1 data classifies incoming delay events into correct root-cause categories, initially on a daily refresh cadence and tightened only where the use case justifies a workaround.
Begin feeding data into the maintenance datamart.

Phase 3 (H2, Months 9-12): Roll across remaining sites including Burns Harbor (HSM delay pattern analysis — Pareto on delay codes as the "super simple" first step per Miles B's framing). Cross-site benchmarking on maintenance effectiveness metrics.

Draft acceptance criteria for blueprinting. Proposed roadmap-stage examples only: linkage precision high enough for maintenance teams to trust sampled matches, enough delay coverage to make the dashboard operationally useful, and demonstrated reduction in misattribution on reviewed events. Illustrative confidence bands can be used during design: high-confidence matches flow into the integrated view, ambiguous matches are quarantined for review, and low-confidence matches are excluded until resolved. Final thresholds are set in the project blueprint after the data assessment.

Entry site: Cleveland — strongest champion density (Jamie Betts, Paul Aaron Dash, Dan Hartman), both Tabware and ops reporting systems accessible, 1SP provides clear production-impact measurement.

Economics. - Value: $2-5M/year per site (misattributed delay correction × production value recovery). Base case: $14M/yr across 4-5 sites. - Investment: $1.84M total program — Phase 1 $540K (Cleveland entry), Phase 2 $684K (Middletown + Tilden extension), Phase 3 $612K (cross-site platform) - Quick win: First misattribution analysis deliverable in 4-6 weeks

Champions: Jamie Betts (Cleveland 1SP Maintenance), Steve Longbottom + John Houston (Middletown), Pete Austin + George Harmon (Tilden), Miles B (Burns Harbor).

PRJ-09: Knowledge Capture / Virtual SME¶

The problem. Across every site, the people who know how the plant actually works are approaching retirement. Three leaders at Middletown independently called this "the biggest problem facing the plant." At Burns Harbor's coke plant, three of five section managers are retirement-age and the 54-year electrical manager's knowledge exists nowhere in any system. At Tilden, Adam Bingham is building knowledge bases on his own initiative because no institutional system exists.

Validated at: Middletown (Brian Benning championed, leadership's #1 at Day 5 readout — scope expanded in real time to cover L0/L1/L2 per department), Tilden (Adam Bingham + 1,200-page HPGR manual + knowledge base already started), Burns Harbor (coke plant knowledge cliff + Indiana Harbor binder-scanning engineer), Cleveland (partial — maintenance knowledge recognized but not prioritized).

Phase 1 (H1, Months 1-4): Knowledge capture pilot at Tilden HPGR — best-scoped pilot across all sites: rich sensor data, 1,200 pages of unread manuals, named champion.

Agentic RAG architecture: an LLM agent (model-agnostic, hosted through Azure AI Foundry) that can query multiple sources and reason across them. Phase 1 tools: approved manuals and CMMS (Tabware) query.
Every response must include source citations grounded in retrieved documents; if the agent cannot ground its answer, it refuses and escalates to a human SME. If manuals and CMMS evidence conflict, the system presents the conflict explicitly and defers to the process owner rather than improvising.
During development, the approved-source list is maintained by the project owner; after handoff, ownership transfers to the process owner. No unsourced procedural guidance, ever.
In parallel, begin coke plant knowledge capture at Burns Harbor (most urgent timeline — retirements imminent).
Knowledge capture sessions use a dual format: guided walkthroughs where the expert narrates while operating or simulating the system (captures procedural knowledge in context), and free-form brain dump sessions where the expert talks about whatever they think is important (captures tacit knowledge they wouldn't volunteer in a structured format). The LLM records and structures both.

Phase 2 (H1-H2, Months 5-8): Expand to Virtual SME framework at Middletown (Brian Benning's expanded scope — process control knowledge per department).

Turn Log intelligence: unsupervised sequence mining on 1.3M Turn Log entries to discover recurring maintenance activity patterns that precede failures — automating Brian's manual scanning approach ("I try to find a pattern: these guys worked on it a couple weeks before it broke").
AI-assisted legacy code documentation: LLM ingestion of Fortran 77, CRISP, and PHP codebases to produce function-by-function documentation and flowcharts, replacing the 8-month manual effort and de-risking key-person dependencies (Bruce, 70+, pickle line).
Progressive tool unlocking: add Pi historian query to the agent's toolset once the data platform delivers live feeds. Scale HPGR model to other Tilden equipment.
At Tilden: USW contract RAG deployment — supervisors ask questions, system retrieves relevant clauses with citations, scalable to all USW sites.

Phase 3 (H2, Months 9-12): Cross-site knowledge base. Add sensor anomaly context to the agent's toolset. Virtual SME available to technicians across the footprint.

Draft acceptance criteria for blueprinting. Proposed roadmap-stage examples only: high grounded-answer rate on sampled questions, strong SME acceptance on reviewed answers, clear refusal behavior when evidence is missing or conflicting, and measurable reduction in technician search or onboarding time. Final thresholds are set in the project blueprint.

Entry sites: Tilden (HPGR pilot — best scoped, Adam Bingham champion) and Burns Harbor (coke plant — most urgent, retirement timeline).

Economics. - Value: $1-5M/site (training time reduction, faster onboarding, reduced expert-dependency risk, incident prevention). Base case: $12M/yr across 5 sites. - Risk mitigation: Incalculable — the coal blend models are already lost. Each retirement without capture is permanent knowledge destruction. - Investment: $2.35M total program — Phase 1 $684K (HPGR pilot + coke plant capture), Phase 2 $1,218K (Virtual SME framework), Phase 3 $450K (cross-site knowledge platform) - Quick win: HPGR knowledge base operational in 6-8 weeks

Champions: Adam Bingham (Tilden — grassroots, already using Copilot), Brian Benning (Middletown — process control), Coke Plant Division Manager (Burns Harbor), Lynn Casco + Brad Koski (Tilden mining knowledge).

PRJ-06: Maintenance Workflow & Procurement¶

The problem. At Cleveland, a $500 spare part requires weeks of multi-level PO approval. Requisitions are silently cancelled after 60 days. At Middletown, $104 million in spare parts inventory sits across warehouses with ~10% duplicate items and lead times defaulted to 15 days regardless of actual supply chain reality. At Burns Harbor, 19,000 parts across 6 warehouses with items sitting for 20+ years.

Validated at: Cleveland (procurement bottleneck validated by every maintenance stakeholder), Middletown (Sean championed, $104M inventory quantified, Dave positioned as self-funding starter), Tilden (warehouse with no barcode scanning), Burns Harbor ($63M inventory, John Sabo's corporate-level perspective).

Phase 1 (H1, Months 1-4): Procurement fast-track automation at entry site.

Rule engine handles routine cases, but the default H1 operating mode is approval-based rather than unconstrained auto-execution; tighter automation remains case by case.
ML anomaly detection flags unusual patterns: sudden spikes in orders for one part, price outliers, unusual requester behavior. Humans review flagged items, and low-risk routine cases can be approval-routed with lighter touch controls.
Inventory deduplication analysis uses a hybrid approach: fuzzy string matching on part numbers, embedding similarity on free-text descriptions (catches semantically identical items described differently), and structured attribute comparison on dimensions and material specs.
At Cleveland: critical spares identification (Brian Thompson's 500-part effort) is accelerated with LLM-based extraction from equipment manuals and BOMs to auto-populate item master fields, plus similarity matching to find duplicates already in the system under different names.
At Middletown: lead time inference uses historical PO mining — extracting actual PO-to-receipt timestamps from Oracle transaction history to replace the 15-day default that applies to all 32K parts regardless of actual supply chain reality.

Phase 2 (H1-H2, Months 5-8): Cross-site item master rationalization using the hybrid deduplication at scale across Middletown ($104M), Burns Harbor ($63M, 19K parts across 6 warehouses), and Tilden.

At Tilden: OEM parts catalog auto-import — LLM extracts part numbers and specs from vendor documentation, structured matching deduplicates against existing ELLIPS stock codes. Addresses the root cause of data quality degradation every time new equipment is purchased.
ML demand forecasting per part (production schedule, maintenance plans, seasonal patterns) to compute optimal min/max reorder points, replacing static thresholds set once and never updated.
At Burns Harbor: directly addresses John Sabo's corporate-level perspective on 20+ year dead stock and obsolete auto-reorders.

Phase 3 (H2, Months 9-14): Integrated maintenance copilot — text-based mobile interface first (tablet/phone for work capture, automated work order creation, procedure lookup).

At Middletown: includes building a lightweight CMMS from scratch for the recently absorbed fleet (2.5 years with no CMMS tracking, whiteboard + paper today).
Voice-based capture is a future enhancement added once the text-based NLP pipeline is validated, avoiding the added complexity of automatic speech recognition in noisy plant environments.
Wi-Fi coverage assessment required at Cleveland — connectivity dead zones are a genuine blocker for plant-floor deployment.

Draft acceptance criteria for blueprinting. Proposed roadmap-stage examples only: procurement cycle-time reduction for routine cases, sampled human-review precision on item deduplication candidates, acceptable anomaly catch rate without overwhelming reviewers, and text-first work capture adoption before voice is considered. Final thresholds are set in the project blueprint.

Entry site: Middletown — strongest inventory data ($104M quantified), named champion (Sean), Dave's "self-funding starter" positioning means early returns directly fund the next project.

Economics. - Value: $2-19M/site (procurement cycle reduction + inventory optimization + copilot efficiency). Base case: $23M/yr across 4 sites. - Inventory reduction: 5-10% of $104M (Middletown) = $5.2-10.4M freed working capital - Investment: $2.48M total program — Phase 1 $612K (procurement fast-track at Middletown), Phase 2 $648K (cross-site inventory rationalization), Phase 3 $1,218K (maintenance copilot) - Quick win: Procurement rule-based fast-track operational in 4-6 weeks

Champions: Sean (Middletown inventory), Dan Hartman (Cleveland HSM), John Sabo (Burns Harbor cataloging), Adam Bingham (Tilden warehouse).

Quality & Inspection¶

Connecting quality data across the process chain to catch problems earlier and disposition faster.

PRJ-04: Through-Process Quality & Yield¶

The problem. A chemistry issue at the degasser may not surface until six process steps later at coating inspection. When a quality hold is triggered, the investigation starts from scratch — no linkage between process data and the defect. Middletown's Ametek surface inspection cameras classify at only 60% accuracy on the defect types that matter most (lamination vs. gouge). Burns Harbor's quality disposition system flags coils that nobody reviews until the next day — "80% could be programmed in."

Validated at: Middletown (Chuck championed — 60% Ametek accuracy, 1% quality loss doubled from historical, 100 quality holds/day with 50+ types), Burns Harbor (longest quality chain from coke to plate, automated disposition opportunity), Cleveland (partial — caster segments, but quality not primary pain).

Phase 1 (H2, Months 7-10): Ametek classifier retraining at Middletown — measurable accuracy improvement on the defect types that matter most (lamination vs. gouge).

Approach depends on what the Ametek system exposes (raw images, feature vectors, classification confidence scores, retraining hooks). The two local engineers maintaining the classifiers are the key contacts to map integration options: fine-tuning within the vendor framework vs. building a parallel classification pipeline.
In parallel, automated quality disposition rules at Burns Harbor: encode the quality group's disposition logic (temperature maps, gauge data, coiling conditions, chemistry checked against customer-specific tolerance tables). The quality team says "80% could be programmed in" — auditable rule engine for deterministic cases, ML reserved for the remaining 20% of edge cases.
Also in parallel, SPC modernization at Middletown: ML-based drift detection (CUSUM, PELT, multivariate process monitoring) replacing the retired SAS programmer's static PDF pipeline, catching within-spec drift that classical SPC misses — quantifying the experienced metallurgists' "Spidey sense" across 16-17K tensile tests per month.

Phase 2 (H2, Months 11-14): Through-process quality linkage at Middletown — heat-to-coil genealogy connecting chemistry, process parameters, and surface inspection across the finishing chain.

Heat number propagation as the universal key across L2 systems (BF, BOF, Caster, HSM).
Defect probability classifier: supervised model trained on historical slabs with known downstream outcomes, predicting the probability of defects at each downstream finishing step given caster conditions. Explainable (SHAP values show which caster parameters drove the alert) and actionable (quality team can divert problematic slabs before they consume 6+ process steps of finishing capacity).

Phase 3 (H2-H3, Months 15-20): Cross-site quality benchmarking.

Extend to Burns Harbor's plate mill quality chain (longest process chain in CLF).
HSM rolling model reverse-engineering at Middletown: use AI to understand the Siemens L2 black-box model through input-output mapping and sensitivity analysis before building a replacement — gives CLF full ownership of the resulting model.

Entry site: Middletown — Ametek cameras installed, Chuck is a committed champion, R&D momentum exists (Matt's team already building BOF endpoint models). Ametek API assessment is the critical first step.

Economics. - Value: $15-43M aggregate across 3 steel sites. Base case: $29M/yr. - Ametek improvement: 60% → 85%+ accuracy on critical defects = significant reduction in misclassification-driven quality losses - Automated disposition: 80% of holds auto-resolved = faster coil release, reduced inventory dwell - Investment: $2.70M total program — Phase 1 $756K (Ametek + disposition + SPC), Phase 2 $1,260K (through-process quality), Phase 3 $684K (cross-site quality + L2 reverse engineering)

Champions: Chuck (Middletown quality), Matt + Eric Bridge + Eric Welty (R&D team), Senior Ops Leader (Burns Harbor quality).

PRJ-05: Cobble & Process Risk Prediction¶

The problem. Cobbles at the HSM are the single most disruptive operational event — production stops, equipment may be damaged, and the root cause investigation rarely produces systemic learning. BF stove management at Burns Harbor is single-expert dependent. BF burden mix optimization at Middletown relies on engineering judgment rather than data-driven models.

Validated at: Cleveland (cobble frequency highest in CLF, Dan Hartman championed), Middletown (R&D actively working on cobble prediction, IBA data available), Burns Harbor (6 BFs across CLF, BF Process Engineer has Bethlehem Steel thermal model heritage, BF stove dependency risk).

Phase 1 (H2, Months 7-10): BF stove optimization at Burns Harbor — bounded problem, BF Process Engineer is the most technically sophisticated champion across all sites.

Augments his existing Bethlehem Steel thermal model heritage with physics-informed ML (first-principles heat transfer equations combined with data-driven residual learning for refractory degradation, uneven heating, and charging variability). Sensor limitations (2 of 4 pyrometers working) make the physics-informed approach necessary — pure data-driven models would be too fragile with current instrumentation.
At Middletown: stove tender decision support takes a different approach — expert system / rules capture encoding the stove tender's decision logic as transparent if-then rules, critical for operator trust in a BF environment.
Sinter plant optimization at Burns Harbor deferred until BF Process Engineer endorses the problem definition — his 4-month AI vendor trial found zero incremental value, so sinter must be framed as a genuine gap his existing Height and Heat thermal model doesn't cover.

Phase 2 (H2, Months 11-14): Cobble prediction at Cleveland HSM.

Two-stage architecture: pre-roll slab features (temperature profile, chemistry, gauge at entry) produce a risk score, then real-time stand-level signals (roll force, motor current, vibration, tension per stand) detect instability during rolling.
Same architecture at Middletown but with richer in-roll features from IBA's millisecond-level data.
Explicitly positioned as predictive analytics ("detect risk conditions sooner"), not real-time control — distancing from the California startup's 6-month failure at Burns Harbor and ArcelorMittal's 2.5-year zero-result caster project.
At Burns Harbor: cobble prediction deferred until BOF off-chemistry (PRJ-08) and plate hit list (PRJ-07) demonstrate that AI delivers — credibility must precede the conversation about past failures.

Phase 3 (H2-H3, Months 15-20): Cross-site BF optimization covering 6 blast furnaces.

BF burden mix optimization at Middletown (IH7 as starting point).
Context-aware alarming across sites: hybrid of rules-based thresholds for known failure modes plus ML envelope (unsupervised anomaly detection) for subtle deviations that depend on product/grade/speed combinations.

Entry site: Burns Harbor (BF stove optimization) — Palmer named BF stoves explicitly. BF Process Engineer has deep domain expertise and existing models.

Economics. - Value: $8-35M across sites (cobble prevention, BF thermal optimization, burden mix). Base case: $22M/yr. - BF stove: Single-expert dependency removed + thermal efficiency gain - Cobble: Each prevented cobble avoids 4+ minutes of production loss plus potential equipment damage - Investment: $2.63M total program — Phase 1 $648K (BF stove optimization + MDT decision support), Phase 2 $1,302K (cobble prediction at Cleveland + MDT), Phase 3 $684K (cross-site BF + alarming)

Champions: BF Process Engineer (Burns Harbor), Dan Hartman (Cleveland cobble), Matt + R&D team (Middletown cobble), Bill (Burns Harbor HMI/PA).

Asset Performance¶

Moving from reactive to predictive maintenance, starting with high-value proof points.

PRJ-03: Predictive Maintenance Platform¶

The problem. Reactive maintenance rates range from 70% (Cleveland) to 30% (Tilden fleet). Even at sites with better PM programs, predictive capability is minimal — vibration data exists but is analyzed manually, condition monitoring is route-based rather than continuous, and no systematic correlation between process data and equipment failure patterns. At Cleveland, fragmented condition monitoring (AssetWatch at powerhouses only, cloud vibration sensors on 1SP cranes that plant IT didn't know existed, Viz route-based readings with selective coverage) means assets generate data nobody watches.

Validated at: Cleveland (PdM PoV greenlit by Chad Asgaard — Mar 16: "get started right away"), Tilden (fleet PdM + HPGR pilot, Pete Austin championed fleet lifecycle models), Burns Harbor (belt system PdM + multi-asset opportunities, BF alarm triage across 100+ HMI screens), Middletown (fleet vehicles recently absorbed, 2.5 years with no CMMS tracking).

Current status: ACTIVE. Cleveland PdM PoV proposal v3 submitted. Eight-week multi-asset charter covering three target assets at 1SP plus a multi-asset data readiness assessment. Contract in final approval. Separate SOW from this engagement.

Phase 1 (H1, Weeks 1-8): Cleveland PdM Proof of Value. The PoV is not a single-asset experiment — it is the first working section of an integrated 1SP health monitoring system. A tiered multi-asset approach spanning the data-richness spectrum:

(1) BOF Bag House ID Fans — highest data readiness, 4 fans, rich differential pressure and temperature data, multimodal anomaly detection combining vibration with process parameters, environmental compliance threshold.
(2) BOF Scrubbing System — predictable 3-week degradation cycle, -5 heats/day when degraded, RUL prediction modeling the degradation trajectory.
(3) Crane 300 — single point of failure for 1SP, near-zero instrumentation, MCSA from substation data, instrumentation gap assessment as core deliverable.
Plus a multi-asset data readiness scorecard for every asset reachable through 1SP subsystems, a system design wireframe (Week 8 demo), and an IE-led data foundation blueprint (L0-L1 to cloud pipeline, asset onboarding playbook).

Phase 2 (H1-H2, Months 3-10): Expand Cleveland PdM to additional assets and launch Tilden fleet PdM.

Cleveland expansion: hot metal cranes, HSM work rolls, descale pumps, caster segments.
Tilden fleet PdM: per-tire survival analysis (Cox proportional hazards or ML survival model) combining position, duty cycle (tonnage hauled, road grade, distance), environmental conditions, and tire metadata, not hours-based. Addresses Chad's tire skepticism by focusing on operational patterns rather than component lifespan.
Wheel motor rebuild optimization ($300K/unit, 4th rebuild decision point).
Each site gets independently trained models — the common investment is in the pipeline and infrastructure, not the models.

Phase 3 (H2, Months 11-18): Cross-site PdM platform.

Scale to Burns Harbor belt system and BF alarm analytics (frequency/pattern analysis on structured HMI alarm logs to identify alarm clusters before known failures, nuisance alarm masking, and alarm flood patterns).
Integrate with maintenance datamart (PRJ-01) for cross-site failure pattern recognition.

Draft acceptance criteria for blueprinting. Proposed roadmap-stage examples only: useful lead time before failure or maintenance intervention, false-alert rate low enough for planner adoption, evidence that alerts influence maintenance decisions, and measurable reduction in downtime or avoided cleaning/failure events. Final thresholds are set in the project blueprint.

Entry site: Cleveland — PdM PoV in final contract approval under separate SOW.

Economics. - Value: $3-24M/site depending on asset base (steel sites: $3-12M from downtime reduction; Tilden fleet: $9-24M including tire lifecycle, wheel motor rebuild optimization, and fleet availability). Base case: $36M/yr across 4 sites. - Cleveland PoV: Multi-asset proof of value delivering results within 8-week charter — foundation + 3 deep-dive assets. Per-asset cost drops 60-70% in Phase 2 because foundation exists. - Investment: $2.08M total program — Phase 1 $236K (Cleveland PoV, separate SOW, firm), Phase 2 $1,302K (Cleveland expansion + Tilden fleet), Phase 3 $540K (cross-site PdM platform) - Quick win: Cleveland PoV Week 8 demo — 1SP health monitoring system with live analytics on 3 assets

Champions: Paul Aaron Dash + Jamie Betts + John Messi + Brian Thompson (Cleveland 1SP), Pete Austin + Chase Lincoln (Tilden fleet), Adam Bingham (Tilden HPGR), BF Process Engineer + Bill Barker (Burns Harbor).

Plant Logistics¶

Optimizing the movement of materials through and between facilities.

PRJ-07: Intra-Plant Logistics Optimization¶

The problem. Burns Harbor's GM identified shipping velocity as the number one priority. Six people manage 220,000+ tons per month on a 1980s IMS coil tracking system. When inventory exceeds 135,000 tons, the plant stops making steel because there is nowhere to put coils. At Middletown, 40 coil loads per day require 2 hours of manual planning per person. At Tilden, train scheduling between the mine, concentrator, and dock is done manually.

Validated at: Burns Harbor (GM's #1 priority, $22-60M coil velocity opportunity, strongest signal), Middletown (Chris + West championed, Palmer's #1 named priority), Tilden (Kevin championed train scheduling, Ryan ranked logistics #3), Cleveland (partial — slab logistics identified).

Phase 1 (H1-H2, Months 3-6): Plate mill shipping hit list automation at Burns Harbor — goes beyond generating a list: trigger actions automatically for low-risk items (auto-combine partial rail cars, auto-assign met releases for clear-cut cases, auto-flag priority changes).

Dave's vision: "that should be a process, not a meeting." The 10-15 daily action items should execute themselves where the logic is deterministic, with human review reserved for exceptions. Achievable in weeks because Dave already built the Power BI infrastructure and the team knows the decision logic.
In parallel at Middletown: door status system first (pure visibility — departments enter door availability windows, dispatchers see all door statuses on one screen, eliminates "truck arrives and discovers the door is down").
Then coil movement optimization (40 loads/day routing and scheduling) once baseline movement patterns are established through GPS data from MobileCom.

Phase 2 (H2, Months 7-12): Full coil velocity optimization at Burns Harbor — birth-to-ship tracking, automated scheduling, inventory threshold management.

Predictive alerts when approaching the 135,000-ton production-stop threshold.
Quality disposition automation (PRJ-04) is the key that unlocks shipping velocity — every hour of disposition delay cascades into 4-5 additional handling steps per misrouted coil.

Phase 3 (H2, Months 13-18): Cross-site logistics.

Tilden mine-to-dock scheduling: constraint satisfaction model (vessel arrival windows from BCS, crew availability, dock capacity, weather windows) replacing Kevin's 3-4 hour manual replanning. Optimization layer added once the constraint model is validated.
Cleveland slab movement (OR-based slotting, progressive complexity from rule-based heuristics to optimization).

Entry sites: Burns Harbor (GM's #1, highest volume) and Middletown (Palmer's #1, coil planning validated).

Economics. - Value: $22-60M aggregate (Burns Harbor shipping velocity dominates: additional tons shipped × margin + cycle time reduction + reprocessing avoidance). Base case: $41M/yr. - Middletown: $2-5M (planning time reduction + coil handling optimization) - Tilden: $2-6M (train scheduling, dock utilization) - Investment: $2.59M total program — Phase 1 $684K (plate hit list at BH + coil planning at MDT), Phase 2 $1,218K (BH coil velocity), Phase 3 $684K (Tilden scheduling + Cleveland slab)

Champions: Sam + Paul (Burns Harbor shipping), Dave Holter (Burns Harbor plate), Chris + West (Middletown), Kevin (Tilden train scheduling).

Process Optimization¶

Using data to optimize process chemistry, steelmaking parameters, and production control.

PRJ-10: Process Chemistry Optimization (Tilden Concentrator)¶

The problem. Tilden's concentrator, designed in 1974, processes ore that has fundamentally changed over fifty years of open-pit mining. The plant spends $50 million per year on chemical reagents and recovers approximately 70% of iron content — below the plant's design benchmark of ~75% and well below the 80% achievable with optimized hematite flotation. There is no predictive link between ore variability and concentrator response. Met techs dose dispersant (PAA) based on experience and subjective beaker tests — section-to-section and operator-to-operator variability is unmeasured. Weight recovery ranges 36-42%, and Ryan Korpela estimates 0.5% weight recovery improvement equals 100,000 tons, worth tens of millions per year.

Chad Asgaard directed on March 24: "The biggest opportunity in Cleveland [Tilden] is how we operate the concentrator and making contents the most efficiently... I want to focus high on process optimization."

Validated at: Tilden (strongest single-site business case in the portfolio — $8-18M/yr, Todd Davis + Dan Collins + Sean Halston championed, Ryan ranked mine-to-mill as #1 priority, Keith Holmgren confirmed as primary SME). Steel analog exists at Burns Harbor (coke plant chemistry) and across BFs (burden mix).

Current status: ACTIVE. Project charter v3 submitted. Phase 1 (Proof of Value) scoped and priced at $312,000 firm. Phase 2 (Full System Build) scoped at $1.3-1.56M indicative. Erico Lemos managing as project lead. Pending contract execution and data readiness confirmation.

Phase 1: Proof of Value (Weeks 0-7, $312K firm). A lean 3-person Vooban team proves the core thesis: do existing process signals correlate with recovery outcomes strongly enough to justify the full system? Delivers:

(1) Baseline recovery variability model characterizing section-to-section and operator-to-operator performance.
(2) Retrospective desliming optimization model — a supervised regression model trained on historical process variables (tailing sump levels, DTU pump speeds, thickener profiles, met balance outcomes) predicting optimal PAA dose per section, validated by backtesting on held-out historical periods including ore variability events.
(3) Preliminary flotation recovery correlation linking desliming optimization to downstream stability.
(4) Feature importance ranking and signal strength assessment.
(5) Evidence-based go/no-go recommendation for Phase 2.
Phase 1 stands on its own even if Phase 2 does not proceed — for $312K, Cleveland-Cliffs eliminates the uncertainty about whether the $1M+ Phase 2 investment will pay off.

Phase 2: Full System Build (Weeks 8-19, $1.3-1.56M indicative). Proceeds only after the Week 7 gate confirms Phase 1 success. The full Vooban team (7-8 FTE) and IE team (2.25-3 FTE) build the complete system:

(1) TLD-54 Beaker Test Vision — camera-based settling measurement at PSI stations replacing subjective visual assessment, starting with classical image processing (edge detection for settling interface height), with CNN-based classification as a future enhancement.
(2) TLD-32 Live CRP Engine — real-time process state evaluation with rule-based bottleneck identification (encoding Keith's CRP framework) plus LLM-generated recommendations drawing on captured process knowledge.
(3) Operator advisory dashboard with per-section dispersant guidance, shift handoff views, and met tech feedback tracking.
(4) Flotation feedback loop validating desliming optimization impact on downstream recovery.
The ML layer generates recommended G2 setpoint and membership-function adjustments but operates in advisory mode — G2 still runs the control loop, operators see a smarter G2, not a replacement.

Phase 3: Mine-to-Mill Optimization (Roadmap, Months 11-18). Extends optimization upstream.

Geostatistical interpolation (kriging/IDW) on drill assay data builds a 3D ore body composition model, then material tracking follows flow from blast to stockpile to concentrator feed.
Blast pattern optimization feeds downstream into the concentrator response models built in Phase 1-2.
Pellet plant parameter integration extends the optimization chain through to final product quality.

Entry site: Tilden — only mine in CLF, $50M reagent spend, Chad's explicit directive.

Economics. - Value: $8-18M/yr addressable at full deployment (reagent reduction 5-10% = $2.5-5M + recovery improvement toward 75-80% at ~$7.5M per percentage point = $3.5-7.5M + throughput stability $2.5-3M + beaker vision measurement consistency + operator variance reduction) - Phase 1 standalone value: $0.5-2M (elimination of uncertainty, actionable analytical outputs, data quality map, quantified opportunity) - Strongest single-site value in the entire portfolio - Investment: $2.26-2.52M total program — Phase 1 $312K (firm), Phase 2 $1.3-1.56M (indicative charter), Phase 3 $648K (template estimate) - Quick win: Phase 1 baseline recovery model and feature importance ranking in 7 weeks

Draft acceptance criteria for blueprinting. Phase 1 gate: do existing process signals correlate with recovery outcomes strongly enough to justify the full system investment? Phase 2: grounded-answer rate on recommendations, operator acceptance of advisory setpoint changes, measurable movement in recovery variability and reagent efficiency during advisory use. Final thresholds set in the project blueprint.

Champions: Keith Holmgren (Sr Director Mining Technology, primary SME), Todd Davis (Lead Process Engineer, Grinding), Dan McGrath (Process Engineer, backup SME), Sean Halston (ECSM Concentrator Ops). Ryan Korpela (GM) as executive sponsor. Note: readout feedback indicated need for mine-side engineering champion — to be identified.

PRJ-08: Caster Chemistry & Steelmaking Optimization¶

The problem. At Burns Harbor, 5% of heats are off-chemistry (carbon and sulfur contribute 3%), each representing approximately $1 million in exposure on a 300-ton heat. Caster plugging averages 25 events year-to-date, each costing 80 minutes of production plus $40,000 in tundish replacement. The process data exists — SQL records going back to 2001, PA team confirming the data is "pretty kind of ready to feed into an LLM."

Validated at: Burns Harbor (Dave Holter's #1 priority — 20 years of chemistry data, in-house models, clear ask: "start with something super simple"), Middletown (R&D building BOF endpoint model with Copilot), Cleveland (partial — caster segments).

Phase 1 (H2, Months 7-10): Off-chemistry analysis at Burns Harbor — Dave Holter's 20-year dataset, SQL since 2001. Three-phase approach per Dave's "start with something super simple" directive:

(1) Descriptive Pareto — which categories of off-chemistry are most frequent, by shift, operator, grade, BOF unit. Answers Dave's morning review questions with data instead of manual spreadsheets, achievable in weeks.
(2) Root cause classification — categorize each off-chem event as man (operator deviation from model), machine (L2 model error), or process (material variability) using structured attribution on the 20+ years of SQL data.
(3) Correlation analysis — statistical/ML analysis of which upstream variables (scrap mix, temperature, blowing parameters) predict off-chemistry outcomes. Each phase builds on the prior, progressively deeper.
At Middletown: Vooban's role is advisory + productionize — advise R&D on ML best practices during their in-flight BOF endpoint model development, then productionize once they have a validated model. Supporting their engineers, not replacing them.

Phase 2 (H2, Months 11-14): Caster plugging prediction at Burns Harbor — historical pattern analysis correlating 25+ plugging events with grade, tundish heat sequence, secondary processing conditions, and timing variables.

Explicitly positioned as "predictive analytics to fly earlier" (detect risk conditions sooner), not real-time control. Distance from the ArcelorMittal approach which likely attempted real-time intervention. The institutional memory of that 2.5-year failure means framing matters as much as technical approach.
Also investigate Cleveland's Prime Metals model: if the failure was adoption/integration (most likely), the approach shifts from rebuilding the model to wrapping it in a usable UX layer and integrating into L2/operator workflow.

Phase 3 (H2-H3, Months 15-20): Cross-site caster reliability analytics (Middletown as benchmark — Matt's weekly cross-site meetings). Statistical factor analysis identifying which operational variables (practices, equipment age, staffing, maintenance investment, product mix) statistically differentiate high-performing sites (Middletown) from low performers — answering Matt's question: what does Middletown do differently?

Entry site: Burns Harbor — Dave Holter is the strongest individual champion across all sites. Data is accessible and well-structured. PA team is an ally.

Economics. - Value: $11-33M (off-chemistry: 5% × heat value × improvement rate + caster plugging: 25 events × production loss × prevention rate). Base case: $22M/yr. - Investment: $1.42M total program — Phase 1 $390K (off-chemistry analysis), Phase 2 $576K (caster plugging prediction), Phase 3 $450K (cross-site caster analytics) - Quick win: Off-chemistry pattern analysis deliverable in 6-8 weeks

Champions: Dave Holter (Burns Harbor — "Steel Dave"), Doug Fortner (PA Steelmaking, SQL data owner), Matt + R&D team (Middletown cross-site caster analytics).

PRJ-11: Coke Plant Operations & Battery Vision¶

The problem. Burns Harbor's coke plant operates two batteries (164 ovens) on approximately 19-hour coking cycles with manual heating control. The coal blend optimization models were built by PhDs who were let go approximately 18 months ago — nobody knows where the models are. Three of five section managers are at retirement age. Data is scattered across 12+ locations. Environmental compliance is existential (no desulfurization facility).

Validated at: Burns Harbor (Coke Plant Division Manager championed — electrical engineer background, automation mindset, already building Battery Vision components). Scalable to 3 additional CLF coke plants.

Phase 1 (H1, Months 1-4): Knowledge capture + data consolidation.

Collect scattered coke plant data (12+ locations) into a unified view — Bill Barker's iFix dashboard integration ("can deliver in hours/days") is purely data integration and visualization, no ML needed for the dashboard itself.
Document critical tribal knowledge before retirements using guided walkthrough and free-form capture sessions.
Coke plant delay classification: zero-shot LLM classification on Mike Zamuta's freeform delay entries into structured categories without requiring labeled training data — enables Pareto analysis of delay drivers (e.g., "show me all Larry car delays for the past three years").
In parallel, attempt to recover the original coal blend model artifacts (code, notebooks, spreadsheets) built by the PhDs — one of them "maybe" would participate in knowledge capture. Whether or not recovery succeeds, the recovered logic informs feature engineering for Phase 3.

Phase 2 (H1-H2, Months 5-10): Battery Vision digital twin using physics-informed ML.

First-principles thermal models (heat transfer equations for coke ovens) are combined with ML to learn the residual that the physics model cannot capture: refractory degradation, uneven heating, charging variability. The underlying thermodynamics are well understood, but real-world deviations require data-driven correction.
Push scheduling optimization treats the battery as a sequencing problem (constraint satisfaction across 164 ovens, coking time, heating patterns).
Sensor limitations (2 of 4 pyrometers working) make the physics-informed approach necessary: pure data-driven models would be too fragile with current instrumentation.

Phase 3 (H2, Months 11-16): Coal blend optimization rebuild from historical production data (blend recipes, coke quality outcomes, oven conditions).

New models using modern methods, informed by any recovered insights from the original PhD work.
Tom Zenzian (corporate coal buyer) provides domain constraints (8 coal types, VM/sulfur/reflectance/contraction rates, low-sulfur mandate).
Environmental compliance analytics — existential for Burns Harbor (no desulfurization facility).

Entry site: Burns Harbor — only site with this combination of urgency (retirement timeline) and champion readiness.

Economics. - Value: $7-17M/yr (thermal efficiency, push scheduling optimization, blend optimization, environmental compliance). Base case: $12M/yr. - Scalable to 3 additional CLF coke plants - Investment: $2.16M total program — Phase 1 $390K (knowledge capture + data consolidation), Phase 2 $1,260K (Battery Vision digital twin), Phase 3 $510K (coal blend optimization) - Quick win: Tribal knowledge capture operational in 4-6 weeks

Champions: Coke Plant Division Manager (Burns Harbor — electrical engineer, automation background, already building pieces).

Deferred Project¶

PRJ-02: Production Scheduling & S&IOP (Defer to Program Expansion)¶

Why deferred. No H1 entry point. Requires mature data foundation (H1 ops-maintenance + H2 quality data) before scheduling optimization is viable. Score: 1.80. Revisit at Gate 2 (Month 12) when the data foundation supports it.

Value when ready: $10-42M enterprise (cross-site scheduling, S&IOP integration). This is the highest-value H3 project and the natural destination of the program — it represents the fully integrated plant that the earlier horizons build toward.

Portfolio Summary by Horizon¶

Total program: $22.6M over 24 months. Base-case annual value at full deployment: $224M/yr.

Horizon 1 (Months 0-6) — "Bridge the Gap" — $3.5M¶

Project	Entry Site	Ph1 Cost	Quick Win Target	Status
PRJ-01 Ops-Maint Integration	Cleveland	$540K	Misattribution analysis (4-6 wks)	Planned
PRJ-03 PdM PoV	Cleveland	$236K	PoV results (8-wk charter)	Active — contract pending
PRJ-06 Maintenance Workflow	Middletown	$612K	Procurement fast-track (4-6 wks)	Planned
PRJ-07 Logistics	Burns Harbor + MDT	$684K	Plate hit list (4-6 wks)	Planned
PRJ-09 Knowledge Capture	Tilden + Burns Harbor	$684K	HPGR KB (6-8 wks)	Planned
PRJ-10 Concentrator	Tilden	$312K	Recovery baseline (7 wks)	Active — charter submitted
PRJ-11 Coke Plant	Burns Harbor	$390K	Knowledge capture (4-6 wks)	Planned

Horizon 2 (Months 7-12) — "Build the Foundation" — $11.6M¶

Project	Entry Site	Ph1 or Ph2 Cost	Value
PRJ-04 Quality & Yield	Middletown → Burns Harbor	$756K (Ph1) + $1,260K (Ph2)	$15-43M
PRJ-05 Cobble & Process Risk	Burns Harbor (BF) → Cleveland	$648K (Ph1) + $1,302K (Ph2)	$8-35M
PRJ-08 Caster Chemistry	Burns Harbor	$390K (Ph1) + $576K (Ph2)	$11-33M
H1 projects scaling to Phase 2	All sites	$5.3M across 7 projects	Incremental

Horizon 3 (Months 13-24) — "Expand & Optimize" — $7.5M¶

Project	Scope	Ph3 Cost	Value
All 10 projects Phase 3	Cross-site platforms, corporate dashboards	$6.5M total	Incremental
PRJ-04, 05, 08 Phase 2	Full system builds at entry sites	~$3.1M	$34-111M combined
PRJ-02 Production Scheduling	Deferred — revisit at Gate 2	TBD	$10-42M