You Can't Divert What You Never Measured: Industrial Waste Diversion Starts at the Roll-Off

Every industrial diversion number you have ever read is the output of an instrument chain, and almost nobody can describe theirs. I build industrial IoT sensor networks and the real-time monitoring platform on top of them, so my day is spent at the unglamorous end of this: which device produced that reading, when was it last checked against a standard, and what happened to the number between the sensor and the slide. When a deck says 78% diverted from landfill, the question isn't whether 78 is a good score. It's which instrument weighed it, and who checks that instrument.

Here is the reframe, up front. The reported rate is about 10% of the problem; the instrumentation feeding it is the other 90%. A diversion percentage is a derived quantity, three or four transformations downstream of a load cell, a truck scale, a bin count somebody wrote on a clipboard, and a spreadsheet nobody has opened in a year. Every one of those steps can bias the result without throwing a visible error, which is what makes bad measurement worse than none. EPA defines a waste assessment as a systematic review of a site and its operations to identify the quantity and composition of its waste stream. Two measurements. Plenty of plants have neither.

The number everyone cites is wrong

Ask how much industrial waste the US generates and you'll hear 7.6 billion tons. It's in slide decks, trade-press intros, vendor one-pagers. It's also wrong. The figure traces to a 1988 EPA Report to Congress, and it was mostly an estimate of dilute wastewater sitting in on-site surface impoundments in 1985 (measured wet, which is why it dwarfs everything else). It was never a count of the solid byproducts a plant trucks to a landfill.

A 2020 re-analysis in Resources, Conservation and Recycling put actual US non-hazardous industrial waste at roughly 200 to 300 million tonnes a year, with about 100 million tonnes of recoverable material inside that total. That's the same order of magnitude as household trash. EPA's 2018 facts and figures put municipal solid waste generation at 292.4 million tons, of which 146.1 million went to landfill. So industrial and municipal are comparable in scale, and both bury roughly half of what they make.

A stale headline number matters to an operator for one specific reason: it tells you the measurement culture around this stream is loose enough that the sector's most-quoted statistic has been wrong by more than an order of magnitude for close to 40 years, and nobody flinched. If the aggregate is that soft, your plant-level number is softer; far fewer people look at it. A zero-waste-to-landfill badge rewards the claim rather than the weighing, and its arithmetic is rarely held to the standard that applies to the invoice billing you for the same tonnage. We went deeper on how one of those rates collapses under scrutiny in a separate piece on measuring diversion rates you can defend.

What's actually in the roll-off

The container is the right place to instrument, because it's the last point where the stream is one physical thing you can weigh and photograph. The minimum rig is not exotic: a load cell under the roll-off or compactor, and a downward camera over the charge point running a classifier trained on that site's own material. None of that is clever, and it shouldn't be. The point isn't the model. It's ground truth on what the box held, by weight, before anyone turns it into a percentage.

Run that for a few weeks at a typical manufacturing site and the same shape shows up. The largest divertible fraction is rarely the exotic process scrap everyone assumes. It's clean corrugated and stretch film riding into the landfill container because the recycling baler sits further from the presses than the trash compactor does. Distance, not economics. EPA's commercial waste guidance lands in the same place, treating contamination as a signal about convenience rather than about the material. Manufacturing waste reduction gets sold as an equipment purchase. On most floors it's a measurement-and-logistics problem wearing an equipment costume.

Now the part that gets skipped. Instruments have to be trusted, not assumed, and the US has a standard that says so. NIST Handbook 44 sets the specifications and tolerances for weighing and measuring devices, and state weights-and-measures authorities adopt it to cover commercial equipment, the devices used to buy and sell. Your hauler's truck scale sits inside that regime and gets tested by somebody outside your fence. The load cell and bin sensors feeding your sustainability report usually don't, because no money changes hands across them. The number that bills you is held to legal metrology. The number that flatters you is held to nothing.

An uncalibrated input doesn't fail loudly. It biases, in one direction, on every reading, for as long as nobody checks, then launders that bias into a signed report. A scale reading a few percent heavy on baled fiber throws no alarm; it credits tonnage that never existed, consistently enough to pass for a trend line. Sensor drift always wins eventually, so design for it: calibration on a schedule rather than on memory, and reconciliation against an instrument you don't control.

Before I'd trust a plant's diversion figure, I want roughly this much behind it:

1. A calibrated load cell under the container or compactor, logging every haul by weight rather than by truck count, with recalibration on the maintenance schedule.
2. A downward camera over the charge point, grabbing frames often enough to characterize composition. Even a rough classifier beats a purchasing-based estimate.
3. A walked floor path from each press or line to the nearest correct bin. Distance is the quietest cause of contamination there is.
4. Three months of hauler weigh tickets set against your own scale, with any divergence written down rather than averaged away.
5. For every fraction you want to call diverted, a named buyer and the trucking distance to them. No buyer, no diversion. Just storage with better branding.

Where the divertible tonnage actually is

Once you've characterized a stream, the material sorts into three buckets: things with a mature offtake market, things with a market somewhere else, and things you're landfilling no matter what you buy. Industrial byproduct recovery lives almost entirely in the first two, and the split is regional.

Foundry sand is the textbook case. American foundries generate around 10 million tons of spent sand a year, and by EPA's estimate only about 2.6 million get beneficially used outside the foundry, as road subbase, manufactured soil, and cement replacement. Iron, steel, and aluminum sands make up 96% of what does get reused, because their chemistry is predictable enough that a state regulator will sign the reuse. The other seven-ish million tons are landfilled, not because the sand is unusable, but because the plant isn't near a road job or a soil blender that wants it. The same material gets buried in one county and reused in the next, and the deciding factor is a map. Non-hazardous industrial waste sits under RCRA Subtitle D, which leaves the rules to the states, so what counts as divertible is partly a function of which state you're standing in; EPA's Guide for Industrial Waste Management is written for exactly that split.

That regional-market truth is the part vendors skip. A sorter hands you a clean fraction; it can't conjure a buyer for it within trucking distance. No recovery equipment should earn a line in a capex plan without the offtake contract, or at least a letter of intent; a characterized pile with no buyer is a more expensive way to describe your landfill bill. So what exactly are you optimizing when you buy the sorter first?

The fraction with no material market at all is where thermal routes enter: the residual plastics, contaminated fiber, and mixed organics that no blender will touch go to energy recovery rather than a cell. Plants near an established outlet, the kind of regional draw you see around mature waste conversion facilities, have options that an isolated plant doesn't. If you're weighing that residual stream, the equipment side of thermal conversion technology matters less than whether an offtaker exists inside a hundred miles. The metals fraction is its own recovery problem with its own economics, which we walk through in a companion piece on pulling metal out before conversion.

Source separation beats back-end sorting

The uncomfortable part, given that back-end sorting is what I build: the cheapest ton of diversion is the one that never gets commingled in the first place. The optical units, the pick robots, and the whole waste-intelligence software stack are what you reach for after segregation at source has already failed. It's recovery of last resort, and it's priced like it.

Some numbers ground this. EREF's 2024 survey put the average US landfill tipping fee at $62.28/ton, up about 10% in a year and the biggest jump since 2022. That average hides a lot of spread: the Northeast runs north of eighty dollars a ton while the South Central states sit in the mid-forties. For a plant filling several roll-offs a week, the distance between a cheap-tipping region and an expensive one is the entire business case for source separation, or the entire reason reducing landfill waste never pencils out on paper.

Where back-end sorting genuinely is needed, temper the demo. The computer-vision sorting we built for the OWI platform holds better than 98% classification accuracy on live lines, and the honest version is that most of the work was never modeling. It was sensors, lighting, belt behavior, and feedstock condition. I used to think a better network was the answer to a bad sort; it usually isn't. Vendor recall figures are quoted on a demo stream, cleaner and drier than anything an industrial roll-off produces, and optical units that hold high recall on dry material routinely give up double-digit recall once input moisture climbs (water changes what the near-infrared sensor sees, and the ejector acts on what the sensor saw). A model that can't survive a moisture swing hasn't shipped. Most of what gets sold as AI sorting is a rule-based pipeline with a CNN bolted on the front. Precision looks great in the brochure. Recall is where the model quietly lies to you.

None of this scales down cleanly, and the limits are worth naming. Below roughly two roll-offs a week the load cell and camera cost more than the diversion they'll surface, and you're better off with a quarterly hand-sort audit and a clipboard. It also assumes a regional market exists. Where there's no beneficial-use outlet and a cheap landfill next door, characterization won't work as a diversion play; it just tells you precisely what you're stuck burying. And wet, high-contamination, or commingled hazardous streams break every assumption here; those need a manifest and a treatment path, not a diversion strategy. The method earns its keep in the middle: real tonnage, at least one buyer in the region, and a landfill bill big enough to feel. That middle is also where our zero-waste-to-landfill services tend to actually close, because the economics are legible.

The plants that hit their diversion targets aren't the ones that bought the best sorter. They're the ones that put a checked scale and a cheap camera on the container before they bought anything, then believed what it told them. The divertible tonnage is almost never the constraint; the missing measurement is. A reported rate is a claim about an instrument chain. Ask about the chain. If nobody can say when the scale was last checked, you don't have a diversion rate. You have a number.

Sources & Notes

• Municipal tonnage for scale (292.4 million tons generated, 146.1 million landfilled in 2018) comes from EPA's national MSW facts and figures.
• The foundry-sand figures (roughly 10 million tons generated, about 2.6 million beneficially used, 96% from iron, steel, and aluminum) are EPA's, on its page on spent foundry sand reuse.
• Tipping-fee numbers ($62.28 per ton, up 10%, with the regional spread) are from EREF's 2024 tipping-fee analysis.
• For why the 7.6-billion-ton figure is wrong and what the real number is, see the 2020 study Non-hazardous industrial waste in the United States: 100 million tonnes of recoverable resources in Resources, Conservation and Recycling.
• The weights-and-measures standard behind the calibration argument is NIST Handbook 44, adopted by state and local authorities for commercial weighing and measuring equipment.
• The waste-assessment definition is EPA's, from Managing and Reducing Wastes: A Guide for Commercial Buildings.
• Every quantitative claim above is sourced to the public references in this list.