Computer-Vision Scrap Sorting: Cleaner Charge, Less Copper

The phone call that starts this story came from a hot mill, not a scrap yard. A coil of low-carbon strip had cracked along its edges during hot rolling, the surface breaking up into fine fissures that no amount of downstream pickling would hide. The metallurgist already knew the answer before the lab confirmed it. Copper. Somewhere in the charge that fed that heat, an electric motor or a length of wiring harness had ridden a shredded car body straight into the furnace, and a few hundredths of a percent of copper had gone into solution. Once it's in the melt, it stays. You can't oxidise it out the way you skim off silicon or phosphorus, and you can't dilute a finished heat after the fact. The coil was scrap again, this time for real.

That is the quiet tax on steel made from scrap, and it explains why a problem that looks like furnace chemistry actually lives a long way upstream, on a conveyor belt in a recycling plant. Copper is the classic tramp element. It enters as the wiring, motors, and small castings that are bonded to steel in end-of-life products, and shredding liberates most materials but leaves copper clinging to steel in ways a magnet can't undo. As one review of secondary steelmaking puts it plainly, copper and tin cause hot shortness through a loss of ductility and surface defects, and removing copper from the melt, while thermodynamically conceivable, is not something anyone does economically at scale (Jin and Mishra). So the metal carries its contamination forward, heat after heat, unless somebody catches it before it's charged.

The numbers that govern this are unforgiving in their narrowness. Work at the Colorado School of Mines, funded through the U.S. Department of Energy, frames the tolerance band directly: flat products that need a clean drawable surface want copper below roughly 0.06%, while long products like reinforcing bar can live with more than 0.4%, according to that group's accounting of where copper can and can't go. Their stated research goal is to nudge the practical ceiling for new compositions above 0.15 wt% copper, which tells you how tight the current window is. Below copper's melting point the element sits harmlessly, but above it a thin copper-rich liquid film wets the steel's grain boundaries during high-temperature working and the surface tears. That's hot shortness, and it's why a heat aimed at automotive sheet has almost no room for the copper a heat aimed at rebar would shrug off.

Why does this matter more now than it did a generation ago? Because the world is asking scrap to do more. Electric-arc furnaces run on scrap, and scrap-based steel carries a fraction of the carbon of the ore-and-coke route, so every tonne of clean recycled metal is also a tonne of avoided emissions. But the copper that accumulates in the global steel stock doesn't dilute itself. A widely cited analysis in Environmental Science & Technology modelled the copper concentration moving through the steel supply chain and concluded that, with perfectly coordinated trade and extensive dilution, the system can manage tramp copper until around 2050 before the strategy becomes impractical (Daehn, Cabrera Serrenho, and Allwood, 2017). Dilution means buying primary iron units, pig iron or DRI, to bury the copper in clean metal. That works, and it costs money and carbon every time. The cheaper lever is to not let the copper into the heat in the first place.

For decades the scrap yard's tools for that job were blunt and effective only up to a point. An overband magnet pulls ferrous from a mixed stream. An eddy-current separator throws non-ferrous metals off the end of a belt by inducing a repulsive field in conductive particles. Hand-pickers stand over the line and grab what looks wrong. These methods sort by gross physical property, magnetic, conductive, dense, light, and they're genuinely good at separating iron from aluminium from inert fluff. A shredder followed by a magnet and an eddy-current stage will give you a clean ferrous stream and a mixed non-ferrous concentrate, and for a long time that was the state of the art. But the value, and the contamination, hide inside those bulk fractions. What these tools can't do is read a piece of steel and tell you it has a motor welded inside it, or look at a shiny extrusion and know whether it's a high-magnesium 5000-series alloy or a 6000-series structural profile. They also can't tell two alloys of nearly identical density apart, which is precisely the split that decides whether aluminium goes back to extrusion or down to casting. The information that decides charge purity is locked in shape, surface, and chemistry, and a magnet sees none of it.

Teaching the line to see

A camera over the belt stops being a gimmick here and starts being instrumentation. A colour camera, lit properly and running a trained model, can grade scrap by the things that show up visually: the morphology of a fragment, whether it reads as a clean turning or a greasy mixed bale, the presence of an obvious copper winding or a stainless fastener. Recent work has shown deep-learning models classifying and rating steel scrap into the grade categories that mill buyers actually use, the same judgements a veteran inspector makes by eye but applied to every fragment on the belt rather than a sampled handful (Xu and colleagues, 2023). A model doesn't get tired on the eighth hour of a shift, and it logs every call it makes, which turns an inspector's intuition into a data stream you can audit and trend.

The honest limit of a camera is that it sees a surface, and a surface lies. Paint, rust, oxide, oil, and dirt all sit between the lens and the metal that matters. A coated aluminium extrusion and a bare one can look identical to a network that learned on clean samples, and a steel fragment with copper buried inside it presents a perfectly innocent face. Vision tells you what a piece looks like and where it is on the belt, fast enough to track it to a diverter. It does not tell you what the piece is made of. For charge purity, that second question is the one that pays the bills.

Reading composition takes a different kind of eye. Laser-induced breakdown spectroscopy fires a focused pulse at the metal, vaporises a microscopic spot, and reads the light the plasma emits; each element writes its own lines in that spectrum, so the sensor can recover the actual alloy chemistry, magnesium, silicon, copper, zinc, in the moment a fragment passes. Paired with the right classifier the method is precise. One study combining LIBS with a manifold-learning model reported 96.67% accuracy distinguishing five aluminium alloy samples, separating grades that are visually indistinguishable (Harefa and Zhou, 2022). And it works at line speed: a KU Leuven group built a real-time LIBS classifier that sorts post-consumer aluminium into three commercial fractions and returns a decision within milliseconds, fast enough to keep up with an industrial belt (Díaz-Romero and colleagues, 2022). X-ray fluorescence does an analogous job and tends to be the tool of choice where heavy elements and stainless grades dominate.

So the architecture that actually holds up on a floor isn't camera-versus-spectrometer. It's both, each doing what it's good at. The camera handles localisation, throughput, and the gross visual grade, finding every fragment and tracking it down the belt. The spectroscopic sensor interrogates chemistry on the pieces that matter, the ones whose alloy or contamination decides which bin they belong in. The vision system tells the chemistry sensor where to aim and tells the diverter where the piece will be by the time the air jet fires. Fuse the two streams and you get something neither delivers alone: a per-fragment record of what a piece looks like, what it's made of, and where it went. That record is the raw material for a control loop, and it's the part that connects a sorter on the yard floor to the furnace chemistry three buildings away.

From a cleaner belt to a cleaner heat

Charge purity is where the money and the carbon meet. On the aluminium side the prize is keeping wrought alloys out of the cast stream. A clean 6000-series fraction can go back into extrusion; let it mix with castings and high-copper alloys and the whole lot gets downgraded to a tolerant casting alloy, a real loss of value for metal that was perfectly good. The recycling rates leave plenty of headroom to chase. According to the EPA's most recent material-specific figures, aluminium beverage cans were recycled at about 50.4% and aluminium packaging overall at 34.9%, while ferrous metals recovered from durable goods sat near 27.8% (the agency reckons roughly 4.7 million tons of that ferrous fraction was recycled). Better sorting doesn't just lift those rates; it lifts the quality of what gets recovered, which is the part that decides whether recycled metal displaces primary metal or just supplements it.

On the steel side the payoff is the heat that lands on chemistry the first time. If the sorter keeps copper-bearing fragments out of a charge destined for flat product, the melt shop buys less pig iron to dilute, spends less time chasing a heat back into spec, and ships fewer coils that crack at the mill. That's the loop worth closing: the sorter's per-fragment chemistry record feeds the charge make-up model, the furnace result feeds back which scrap grades actually behaved, and the yard learns which suppliers and which shred fractions carry the copper. Running that loop in real time, at the edge, is exactly the kind of problem our edge telemetry and analytics platform is built to carry, because the decisions have to happen in the milliseconds a fragment is airborne, not in a nightly batch report.

None of this is free, and the brochures rarely list the limitations. They are worth listing, because every one of them shows up on a real floor. A spectroscopic sensor reads only the surface it can ablate, so paint, thick oxide, and grease degrade the signal exactly the way they fool the camera; accuracy varies significantly with surface condition, aggressive coatings can drop it hard, and a wet or dusty stream is worse still. Presentation is half the battle. Pieces have to be singulated and reasonably flat so the laser hits metal and the diverter knows where the fragment is, which means belt speed, spreader design, and feed rate matter as much as the model. A sensor pointed at a tangled, overlapping bale won't work no matter how good its classifier is. Calibration drifts, too, and a LIBS or XRF reading is only as trustworthy as the reference standards behind it, which is why traceable calibration, the kind NIST exists to underpin, isn't optional housekeeping but the thing that keeps the chemistry honest week to week. Push the belt faster and detection rates fall; slow it down for accuracy and throughput suffers. There's no setting that maximises both, so the real engineering is choosing the operating point for the grade you're making, and revisiting it when the feed changes.

And the model is never finished. Scrap composition shifts with the season, the supplier, and the commodity cycle; a new shredder feedstock or a change in what the local market is discarding moves the distribution under a model that was trained on last year's belt. Detection quietly degrades, and nothing on the panel announces it. The reject bin fills a little faster, or a little slower, and unless someone is watching the trend the first hard evidence is a cracked coil or an off-chemistry heat. Catching that drift needs a person who understands both the metal and the model, watching reject rates and reviewing the borderline calls, pulling samples to confirm the sensor still agrees with the lab, and retraining before the escapes start landing in heats. The sensor industrialises a sorting decision; it doesn't decide what good looks like. Somebody still has to define the grade boundaries, agree where one alloy fraction ends and the next begins, sample the output, and own the records when a mill auditor asks why a heat went out of spec. None of those jobs disappear when the camera goes in. They change shape, and they move to people who can read what the line is telling them.

What the technology really buys you, then, isn't a magic eye that makes scrap clean. It's a shift in where the decision happens. The old model sorts on bulk physics and discovers the copper problem in the ladle, where it's permanent and expensive. The new model reads shape and chemistry on the belt, where a fragment can still be diverted into the bin where it belongs and the heat it would have ruined never happens. We help plants instrument that handoff, but the metallurgy is older than any of us, and it doesn't negotiate. So the question a yard should ask isn't whether a camera can recognise a piece of scrap. It's whether the line can act on what the sensor saw before the grab closes, because the copper you wave through the gate is copper you've already charged.