Building an intelligence graph for industrial materials.

“I have PET.” “I have PCBs.” “I have aluminum.” These are the sentences with which almost every conversation about industrial waste starts, and none of them is enough to make a decision.

Two loads of PET can be worth very different things. Origin and form matter —bale, flake, pellet— and so do color, contamination level, composition, the process that generated it, its specification, volume, location and the market that needs it. A PCB and an aluminum offcut have other dimensions, but the problem is the same: a category is a label, not a description.

A taxonomy classifies: every material goes in a box. It is useful for ordering, but not for deciding, because the decision depends on relationships no box contains: this material, generated by this process, described by this specification, evaluated by this market, with this outcome.

A material is understood through its relationships. If every stream is represented as a node connected to a site, a process, a specification, a market, an outcome and a time, the system can ask questions a taxonomy cannot answer.

Take clear PET. “PET” does not say whether it is a bale or flake, whether there are labels, caps of another polymer or product residue, whether it comes from a bottling line or from returns, or how much there is each month. With that information missing, the same word describes materials with different destinations, prices and risks. The question stops being “who buys PET?” and becomes “who needs this specification, at this volume, from this location?”.

A PCB rejected on an SMT line is a completely different category, yet the model is the same: identity, process of origin, reason for rejection, batch, specification and destination. We do not assume metal content or values: without analysis, that data is part of what still has to be confirmed. That the same framework works for both is the point.

This is the kind of page where it matters most not to overstate. The graph is not a finished database: it is a knowledge structure built dimension by dimension. Some dimensions are already captured; others are captured in part, with manual support; others are next.

The most defensible data is not what a model believes the material is. It is what actually happened to it: whether it was accepted or rejected, what price was realized, whether there was a quality issue, how much was recovered, what it became, where it went.

A specification alone describes; an outcome alone informs; together they make it possible to compare cases and learn what works. And those are the questions a graph should be able to answer progressively: what happened the last time we saw a similar material, which specification led to a rejection, which destination produced a better outcome, which process generates the most variability.

Relationships matter more than nodes: a list of materials explains nothing. The specification is the dimension with the most explanatory power, because it is the one the market uses to decide. Outcomes are scarce and that is why they are valuable. And every data point needs an explicit level of verification, because mixing the estimated with the measured is the fastest way to lose trust.

More outcomes, more specifications, more relationships, and deeper integration with plant systems. That outcome capture is the subject of Outcome Learning.