Letting the part steer the process.

A cookie recipe is a manufacturing process written for a kitchen.
‍

The inputs are familiar: flour, sugar, butter, eggs, salt, heat, and time. The outcome is easy to recognize. A cookie can be crisp or soft, pale or dark, flat or tall. If you want a crisp edge and a tender center, you adjust the recipe, bake another batch, and see whether you moved closer.

The difficult part is that every parameter interacts with every other one. More butter changes spread. More heat changes browning and moisture. A longer bake changes both. Vary one factor at a time and the quest slows down. Change several at once and the reason for improvement becomes harder to isolate.

Manufacturing engineers do this at a much higher level of difficulty. Injection pressure, melt temperature, cooling time, tool geometry, line speed, thermal profiles, deposition paths, alloy chemistry, and ambient conditions combine to determine the physical result. The number of possible recipes is enormous.

Every process is a terrain of possible outcomes. Optimization is the search for a downhill path through it.

Machine learning incorporates this intuition as a way to train models. Change the parameters, measure how the objective moves, and use that information to choose a better next step. This is known as gradient descent.

Every mature manufacturing process goes through something like this during development, but the loop is slow, manual, and incomplete. Parts leave the machine, samples move to a lab, measurements come later. Engineers decide what to change, often based on experience and gut intuition, while much of the internal structure of the part is never measured at all.

Closing the physical loop
‍

Lumafield’s R&D program has connected that loop end-to-end. A process makes a part, industrial CT records its internal structure, software measures the result and calculates an objective. From there, an optimization system recommends a bounded adjustment, and the process churns out the next part.

Then it happens all over again. We have shown before what CT alone does to process development time, in the die casting case study we published with Nemak, a Tier 1 automotive supplier. AIM continues that shortening, so you can get multiple turns on a physical process in the same day.

The important word is physical. Simulation is indispensable, but every simulation contains assumptions. Everything from material variation to machine and tool wear can make or break the product. The real part registers all of those effects at once.

Industrial CT turns the part itself into feedback: internal porosity, voids, dimensional deviation, material distribution, assembly, and other features that conventional sensors reduce to indirect signals or miss entirely.

We call this Autonomous Intelligent Manufacturing, or AIM.

Four glimpses of the loop

We’ve begun to demonstrate AIM in a few different contexts. These experiments differ in physics, equipment, cycle time, parameters, and quality objectives. But they share a single structure: a system learning from the internal evidence in the things it just made.

Autonomous does not mean unsupervised

‍

Engineers who have begun to implement autonomous systems might understandably have two misgivings: that a factory brain might be free to change whatever it wants, or that the loop would break on an unexpected finding and cause a scramble among its human operators.

That’s where Quality Agent comes into play: Lumafield has developed an intelligent agent that …

Engineers choose what to optimize. They set the permitted parameters and bounds, while equipment interlocks and safety systems remain authoritative. Every iteration produces a physical artifact and a reviewable record, and the system’s authority begins with recommendations and expands only where evidence, validation, and governance support it. That boundary is what makes the idea fit for manufacturing at scale.

The factory as a learning system

‍

Factories have always learned. But until now, their knowledge has been spread across work instructions, control plans, fixtures, process windows, engineers’ brains, and organizational scars.

AIM makes a different kind of learning possible. Each part becomes evidence about the process that produced it. Every single cycle adds a point to the map between parameters and outcomes. The factory searches continuously, within bounds, incorporating direct feedback from the physical world.

That changes more than process tuning. A design system should learn from what factories actually produced: how products varied, where they failed, and which process changes made them better.

Today we are showing controlled lab experiments: solder joints, printed polymers, molded parts, and cast aluminum. The benches are still cluttered, and results are provisional, but they point towards a future in which manufacturing is fast and responsive and the products we deliver push the boundaries of their engineeringlarger implication.

‍

The things we make can tell the machines that make them how to make the next one better.