Inside, it’s a small civilization. Thousands of components made their way through sprawling supply chains before converging on an assembly line where microns make all the difference. The battery layers have to align. Solder joints must form. Adhesives need to fill some gaps while staying out of others. Antennas, cameras, shields, and connectors all have to land in the right place.
While a phone is already intricate enough, that’s just the beginning. A car contains tens of thousands of parts. An aircraft contains millions. The nature of modern manufacturing is such that complexity necessarily disappears: each supplier delivers a component, each station completes an operation, each inspection checks a list of known risks, and a finished object emerges at the other end, hiding all of the steps that went into making it.
The miracle is that this works as often as it does. The danger arises when a failure sneaks in among all that complexity, only popping its head up when the product is in a customer’s hands.
The worst defect isn’t the one that fails a check. It’s the one no one thought to check.
Manufacturers have built formidable systems for measuring known characteristics and addressing failures they expect, but every inspection plan has limitations. By defining what matters in advance, it stops short of what might matter most. A new failure mode slips through because it hasn’t earned its spot on the checklist just yet.
A CAD model denotes intent. Visual inspection records a surface. A test answers the question it set out to ask. Industrial CT records the product exactly as built, inside and out, down to features a few microns across.
For the past seven years, Lumafield has been building industrial X-ray CT systems that allow engineers to look inside the products they’re building without destroying them. Those systems produce three-dimensional images of internal geometry and material distribution: battery layers, solder joints, seals, fibers, voids, cracks, inclusions, and the countless small assembly choices beneath a product’s surface.
One scan reveals a void. But a large population of parts shows whether the void is random, whether it follows a tool, whether it belongs to a supplier, or whether it’s an early, ominous warning that a process is drifting.
This is the dataset behind our foundation model. It’s unique for three reasons:
We’ve trained a model to turn those volumes into useful representations: compressed numerical descriptions that record physical structure and capture the relationships between different samples. No human is naming the features in advance. Instead, the model learns by identifying patterns across a massive collection of scans.
[Caption] Figure 02: This cluster representation gathers scans with useful similarities near one another. Labels are applied after engineers review the physical evidence.
The easiest way to understand this representation is to imagine a map. Every scan is a point somewhere on it. The model places products near one another when their internal construction is similar, and far apart when it’s different.
The model’s map has more dimensions than a person can draw, but two-dimensional projections make some of its structure visible. Batteries with similar layer winding, electrode spacing, weld geometry, and housings form a small city. Connectors with related contact construction amass elsewhere. A unit with different PCB components from all of its peers sits apart.
In this scenario, preconceived ideas of what matters take a back seat. No engineer has to decide beforehand that electrode spacing is the feature worth focusing on. The model preserves relationships that become useful later, when the question comes up.
The model may be responding to any number of part-specific factors. That’s why the real work begins after the pattern appears. At least for now, a human in the loop makes sense of the findings. An engineer returns to the CT volumes, identifies the physical features behind the pattern, connects them to manufacturing context, and tests the hypothesis. The conclusion has to survive contact with the real world.