The Copper Problem: Why Scrap Contamination Is Steel's Hardest Optimization

Copper is steel's most persistent enemy. Unlike carbon or sulfur, it can't be removed once it enters the melt. Every ppm of residual copper degrades downstream mechanical properties — ductility drops, hot shortness appears, and surface quality suffers.

For electric arc furnace (EAF) steelmakers who depend on scrap, this creates an optimization problem with no clean analytical solution. The copper content of incoming scrap is uncertain, blending targets shift with order books, and the cost of over-specifying virgin material erodes margins.

Why Traditional Approaches Fail

Most scrap yards rely on visual grading and supplier history. A trained operator can distinguish HMS 1 from shredded, but neither method gives reliable copper estimates at the charge level. Lab assays arrive too late to change the recipe, and XRF guns — while useful for spot checks — can't characterize a 30-tonne bucket.

The result is a systematic bias toward conservative blending: steelmakers over-purchase low-residual scrap to guarantee spec compliance, paying a premium for certainty they don't always need.

A Data-Driven Alternative

At PRAX, we've built models that learn the relationship between supplier, grade, yard position, and realized melt chemistry across thousands of heats. The system doesn't replace the operator — it gives them a probability distribution over copper outcomes for each candidate recipe, so they can make cost-quality tradeoffs explicitly rather than by instinct.

Early deployments have shown a 6–12% reduction in prime scrap usage without any increase in off-spec heats. The savings compound: less prime scrap means more yard flexibility, shorter procurement cycles, and lower working capital.

What Makes This Hard

The challenge isn't the ML. It's the data infrastructure. Scrap chemistry is observed indirectly (through melt chemistry, lagged by one heat), supplier labels are inconsistent, and yard inventory is tracked manually at most plants. Building the training set requires stitching together ERP records, spectrometer logs, and operator notes — none of which were designed to interoperate.

This is the kind of problem PRAX was built to solve: not a single model, but a system that turns fragmented operational data into reliable, real-time decision support.