ML Raw Material Optimization for Steel: Optimize Your Charge Mix
By Michael Finn on February 22, 2026
Your melt shop superintendent has been blending the charge mix by experience for 22 years. He knows — intuitively — that a 60/40 scrap-to-DRI ratio with a specific grade of pig iron produces acceptable chemistry at a reasonable cost for the structural grades you run most often. But he also knows the scrap yard composition changes every week. Pig iron prices swung 18% last quarter. The DRI silicon content from your new supplier is running 0.3% higher than spec. And the customer just ordered a micro-alloy grade he hasn't made in eight months. His intuition works — until it doesn't. And when it doesn't, the result is a reheat that costs $12,000, an off-spec ladle that gets downgraded $40/ton, or a chemistry miss that requires an alloy addition at the furnace that wipes out the cost advantage of the cheaper scrap he selected. Machine learning doesn't replace your superintendent's 22 years of knowledge — it encodes it, extends it, and runs it against every possible charge combination in seconds. ML raw material optimization models ingest real-time scrap yard inventory, current commodity prices, supplier quality data, target chemistry specifications, and furnace operating constraints to calculate the lowest-cost charge mix that will hit every chemistry target on the first attempt. Not the mix your superintendent thinks will work. The mix that mathematics proves will work — and that saves $8–$15 per ton of liquid steel in raw material costs while reducing chemistry misses by 60–80%.
$8–15
Per-ton savings on liquid steel raw material costs through ML-optimized charge mix selection
78%
Reduction in first-heat chemistry misses when ML selects the charge mix vs. manual blending decisions
340K
Charge mix combinations evaluated per heat by the ML optimizer — a human considers 3–5 at most
$4.7M
Annual raw material savings at a 1.5M-ton EAF mill running ML charge optimization on every heat
What ML Charge Optimization Actually Solves
Manual charge mix selection is a constrained optimization problem that humans solve with heuristics. The melt shop team picks a familiar ratio, adjusts for what's available in the scrap yard, and hopes the chemistry lands in spec. ML solves the same problem mathematically — evaluating every available raw material against every chemistry constraint simultaneously, weighted by real-time cost.
Cost Minimization Under Constraints
Find the lowest-cost combination of available raw materials that satisfies all chemistry targets, furnace capacity limits, and melting constraints — recalculated for every heat based on current prices and inventory.
Chemistry Target Achievement
Hit C, Si, Mn, P, S, Cu, Cr, Ni, Mo, Sn, N, and residual element targets simultaneously — not just the primary elements but the tramp elements that cause quality defects and customer rejections.
Scrap Variability Compensation
Account for the inherent variability of scrap chemistry — the model doesn't assume every bale of #1 HMS is identical. It uses probabilistic composition estimates based on supplier history and source classification.
Alloy Addition Minimization
Reduce expensive ladle alloy additions by front-loading chemistry achievement at the charge stage — getting closer to target before the furnace even taps, instead of correcting in the ladle at 3–5x the cost.
The Optimizer at Work: From Inputs to Recommendation
The ML charge optimizer doesn't operate in a vacuum — it ingests real-time data from multiple plant systems and produces a specific, actionable charge recipe for every heat. Here's the complete data flow from raw inputs through the optimization engine to the operator's screen.
Inputs
Scrap yard inventory by grade & location
Current commodity & scrap prices
Supplier chemistry analysis history
Target grade chemistry specification
Furnace capacity & power profile
Alloy & flux inventory & prices
Evaluates 340K+ combinations in <3 seconds
Outputs
Optimal charge recipe by material & weight
Predicted aim chemistry per element
Total charge cost vs. baseline
Confidence interval for each element
Recommended alloy trim additions
Alternative recipes (2nd, 3rd best)
Steel operations that sign up for ML-integrated operations management connect optimizer outputs directly to charge weigh systems, alloy addition stations, and cost tracking — closing the loop from recommendation to execution to verification.
Charge Mix Scenario Comparison
The real power of ML optimization is visible when you compare the optimizer's recommendation against the default human approach for the same heat. Here's a real-world comparison for a typical structural steel grade on a 120-ton EAF.
Optimize Every Heat. Save on Every Ton. Hit Chemistry Every Time.
OXmaint connects ML charge optimization to your operations platform — linking raw material data, cost tracking, quality results, and furnace performance into a single system that turns every heat into a data point improving the next one.
The optimizer doesn't treat all scrap as equal — it assigns a dynamic quality-cost score to every raw material in inventory based on chemistry reliability, contamination risk, melt yield, and current market price. This score changes in real time as new analysis data arrives and prices fluctuate.
Dynamic Raw Material Scoring Matrix
Material
Chemistry Reliability
Contamination Risk
Melt Yield
Cost Index
ML Score
#1 Heavy Melt
Medium
Medium
92%
$384/t
7.2
Busheling
High
Low
96%
$395/t
8.6
Shredded
Medium
High
89%
$362/t
6.1
Pig Iron
High
Low
99%
$420/t
8.9
DRI / HBI
High
Low
94%
$340/t
9.1
#2 Bundles
Low
High
85%
$310/t
4.3
The Cost-Quality Tradeoff: What the Optimizer Balances
Every charge mix decision is a tradeoff between raw material cost and chemistry confidence. Cheaper materials introduce more variability. Premium materials cost more but hit targets reliably. The ML optimizer finds the optimal point on this curve for every heat — maximizing savings while maintaining the chemistry confidence level your quality standards require. Teams evaluating this optimization capability can book a free demo to see how the cost-quality tradeoff works in practice.
Cost vs. Chemistry Confidence Frontier
Chemistry Confidence
Reject Zone<85% confidence
Risk Zone85–92%
Target Zone92–98%
Premium Zone>98%
Manual avg
ML optimal
Charge Cost ($/ton) →
ML finds the "sweet spot" — higher chemistry confidence at lower cost than manual selection achieves — by exploiting material combinations humans don't consider.
Savings Funnel: Where Every Dollar Comes From
ML charge optimization creates savings through multiple mechanisms — not just cheaper materials. Understanding where the savings come from helps operations teams evaluate the true impact beyond the raw material cost line. Steel operations building their optimization capability can sign up to connect charge optimization data to their cost tracking platform.
Annual Savings Breakdown — 1.5M-ton EAF Mill
$2,100,000
Direct Material Cost Reduction
Lower-cost material combinations that still hit chemistry targets — substitutions humans wouldn't consider because they lack real-time yield and composition data.
$1,200,000
Alloy Addition Reduction
Closer-to-target charge chemistry means less expensive ladle trimming. FeSi, FeMn, and FeCr additions reduced by 30–45% on average.
$680,000
Downgrade & Rework Elimination
Chemistry misses cause product downgrades ($20–$60/ton penalty) or reheats ($12,000+ each). 78% fewer misses eliminates most of this cost.
$480,000
Yield Improvement
Better melt yield from optimized material selection. Less slag, less oxidation loss, better metallic recovery — turning more purchased material into saleable steel.
$240,000
Energy Optimization
Charge mixes that melt more efficiently — better bulk density, appropriate carbon content, optimized DRI ratio — reduce kWh/ton by 3–5%.
Expert Perspective: The Scrap Yard Is Your Biggest Cost Lever
Raw materials are 65–75% of the cost of a ton of liquid steel. Everything else — energy, labor, maintenance, alloys, overhead — fits in the other 25–35%. That means the charge mix decision is the single highest-leverage cost decision made at a steel plant, and it's made 15–30 times per day, every day. Yet most plants make this decision with a spreadsheet, a rulebook, and a superintendent's gut feeling. A 2% improvement in raw material cost efficiency at a million-ton mill is $3–$5 million per year. ML doesn't need to be revolutionary to deliver transformative ROI — it just needs to be slightly better than the human heuristic on every single heat, and those small improvements compound across thousands of heats into millions of dollars. The plants I've seen deploy charge optimization successfully aren't the ones with the fanciest algorithms. They're the ones with the cleanest scrap yard data — accurate inventory, reliable composition analysis, and real-time pricing. The model is easy. The data is the hard part.
Fix Your Scrap Data First
If your scrap yard inventory is inaccurate, your optimizer will recommend materials that aren't there. Invest in scale integration, inventory tracking, and composition analysis before deploying ML.
Close the Loop with Actual Chemistry
Feed every heat's actual tap chemistry back into the model. This is how the optimizer learns which scrap sources really deliver what they claim — and adjusts future recommendations accordingly.
Keep the Superintendent in the Loop
The optimizer should recommend, not dictate. Your experienced melt shop team sees things the model doesn't — moisture in the scrap, an oversized piece that won't fit, a supplier load that looks wrong. Trust the model, but keep humans deciding.
Every Heat Optimized. Every Dollar Tracked. Every Chemistry Hit.
OXmaint connects charge optimization to your complete operations platform — raw material costs, furnace performance, quality results, and maintenance data in one system that makes every heat smarter than the last.
What is ML raw material optimization for steelmaking?
Machine learning raw material optimization uses mathematical models to calculate the lowest-cost combination of available scrap, DRI, pig iron, alloys, and fluxes that will achieve the target chemistry specification for each heat of steel. The model ingests real-time data on scrap yard inventory, current material prices, supplier composition histories, furnace constraints, and grade specifications, then evaluates hundreds of thousands of possible charge combinations to find the optimal blend. Unlike manual charge selection, which relies on heuristics and experience limited to 3–5 familiar combinations, ML optimization considers every available material and every possible ratio simultaneously, finding cost-saving substitutions and combinations that human planners wouldn't identify. The result is a specific charge recipe for each heat that minimizes cost while maximizing the probability of hitting all chemistry targets on the first attempt.
How much can ML charge optimization save a steel plant?
Savings vary by plant size, production route, and current optimization maturity, but typical results range from $3–$15 per ton of liquid steel in total raw material cost reduction. For a 1.5-million-ton EAF mill, this translates to $4.5–$22.5 million annually. Savings come from multiple sources: direct material cost reduction through smarter substitutions (40–45% of total savings), reduced ladle alloy additions from better charge chemistry accuracy (25–30%), eliminated downgrades and reheats from chemistry misses (15–20%), improved melt yield from optimized material selection (10%), and reduced energy consumption from better-melting charge mixes (5%). Most plants achieve payback on the optimization system within 2–4 months of deployment. The ROI is highest at plants with diverse scrap sources, frequent grade changes, and volatile raw material markets.
What data does the charge optimizer need?
The optimizer requires six categories of data. First, scrap yard inventory — what materials are available, in what quantities, and where they're located in the yard. Second, material composition — either from direct analysis (spectroscopy, XRF) or from historical supplier data that provides probabilistic composition estimates by grade and source. Third, current pricing — real-time costs for every raw material including scrap, DRI, pig iron, alloys, and fluxes. Fourth, grade specifications — the chemistry targets and tolerances for the grade being produced on each heat. Fifth, furnace constraints — capacity limits, power profile, tap-to-tap time targets, and operational constraints. Sixth, historical melt data — actual tap chemistry results from previous heats that the model uses to calibrate its predictions and learn how specific material combinations behave in your specific furnace. Data quality is the primary determinant of optimizer performance, particularly scrap inventory accuracy and composition reliability.
Does the ML optimizer replace the melt shop superintendent?
No. The optimizer recommends charge recipes; the melt shop team decides whether to accept, modify, or override. Experienced operators bring contextual knowledge the model doesn't have — visual assessment of scrap quality, awareness of equipment issues that affect melting behavior, judgment about unusual supplier loads, and understanding of downstream processing requirements that may not be captured in the grade specification. The most effective deployments use the optimizer as a decision support tool that handles the mathematical complexity of multi-element, multi-constraint optimization while the human team handles the situational awareness and exception management that requires physical presence and experience. Over time, as the model proves its accuracy, teams typically accept optimizer recommendations on 85–95% of heats and override on the remainder.
How does charge optimization integrate with a steel plant CMMS?
Charge optimization integrates with the CMMS and broader operational platform at multiple points. Raw material inventory data feeds from the yard management system. Pricing data comes from procurement. Furnace performance data — power consumption, tap-to-tap times, refractory condition — comes from the CMMS asset records, because furnace condition affects melt behavior and the model must account for it. Quality results from the lab feed back into the model as training data. Maintenance schedules for furnace relining, electrode changes, and ladle preparation affect the optimizer's scheduling constraints. The CMMS also tracks the cost savings from optimization by comparing predicted vs. actual costs on each heat, generating the ROI data that justifies continued investment. The integration means charge optimization isn't a standalone tool — it's embedded in the operational platform that manages every aspect of the steelmaking process.
