What can AI really do in additive manufacturing?

In additive manufacturing, "artificial intelligence" has become a buzzword. It's advertised at trade shows, featured on magazine covers, and almost never accompanied by the crucial question: What exactly does it do, what data does it use, and what decisions does it make on its own?

This article separates three things that are usually mixed together: what AI already contributes to LPBF, what is under development —including ours— and what no commercial machine does today, no matter how much it is suggested.

Why the LPBF is a particularly good site for this

Powder bed laser melting has three properties that almost no other manufacturing process combines at once.

It generates an enormous amount of data per part. A thirty-hour build consists of thousands of layers, and each layer is an image of the powder bed, an oxygen reading, a pressure reading, several temperatures, and a position reading. None of that needs to be instrumented separately: the process already produces it.

The defects are visual and recurring. A build-up, a bald spot due to lack of powder, a streak from the dispenser, or dirt on the printing area all have a characteristic appearance. An experienced operator can recognize them in a photo. That's exactly the type of problem a trained model handles well.

And the cost of not seeing them is brutal. A flaw in the two-hundredth layer that no one detects until the very end can wipe out thirty hours of machine time, the powder, the argon, and the deadline you gave your client. The value here isn't in optimizing a percentage: it's in not throwing away an entire structure.

What already works

See what's happening, layer by layer

It's the mature application, and the one that's least advertised because it doesn't sound futuristic. A build chamber takes a picture of the powder bed after each coating, and a model classifies what it sees.

In our ALBA models, this model is trained with deliberately generated defective impressions —we'll come back to this, because it's the difficult part— and looks for four specific things:

Around the chamber, the other sensors provide context: redundant oxygen levels, chamber pressure and temperature, laser system and optical path temperature, filter clogging via differential pressure, and powder level via millimeter-wave radar. A re-growth with a clogging filter is not the same as a re-growth with a clean filter.

Reduce the number of experiments to qualify a material

Qualifying a new alloy involves making test specimens, measuring density, cutting, etching the sample, examining it under a microscope, adjusting, and repeating. Each step costs powder, machinery, and weeks.

Here, statistical methods and surrogate models have been helping for years to choose which experiments to run instead of scanning the entire process window. They don't invent parameters: they halve the number of times you have to ask the machine.

With one condition that almost no one mentions: you need to be able to adjust the parameters. If the machine comes with closed and encrypted parameter sets, there's no process window to explore or personal data to accumulate. That's why open parameterization isn't just a checkbox on the technical specifications: it's the prerequisite for any of this to be possible in your shop floor.

Estimate before printing

Knowing how long a build will take and how much it will cost before it even begins is incredibly useful. It's important to clarify what this is and what it isn't: this isn't artificial intelligence, it's simulation. The trajectory is traversed layer by layer, and the results are summed. We explain this because calling a deterministic calculation AI is precisely the kind of thing that makes people doubt what truly is AI.

What's in development, including ours

Detecting is not correcting. The next step, and the one we're at, is for the machine to detect a buildup and act on its own—increase the application time, increase the power of the recirculation pump, and disable the affected part to save the rest of the system.

When it arrives, it will act layer by layer: between one layer and the next. Not within the layer. It's a boring distinction and the most important one in the whole article, so it's in its own section.

What no commercial machine does today

Close the loop on the fusion pool. Controlling the pool geometry in real time while the laser is scanning, not between layers. This is what many people understand when they hear "closed-loop control with AI," and it's not offered by any commercial LPBF machine that we know of today. It exists in labs and in articles. Not in a machine you can buy.

Replacing Inspection. No powder bed monitoring system can replace a CT scan today. It shows the surface of the last layer, not the consolidated interior. They are complementary, not alternatives.

Certification. Monitoring provides evidence for a qualification dossier; it doesn't sign it. Anyone who tells you their AI certifies aerospace parts is telling you something else.

It won't work without your data. A model trained on the alloys and thicknesses of another model won't be accurate with yours without being retrained.

The bottleneck is not the model: it's the data

This is the part that's not in the brochures. Training a defect classifier requires a lot of defect images, and the problem is that a well-tuned machine hardly produces any. Defects are rare by definition, and each one costs a lot of work.

The solution is uncomfortable, but it's the only one: intentionally manufacturing incorrectly. This involves creating a lack of powder, tearing the rubber, dirtying the printing area, and forcing the build-up process. It means wasting machine time and kilos of material generating exactly what you're trying to avoid.

This leads to two practical consequences for the buyer:

• A manufacturer who has trained their actual model can tell you how it generated the defects. If they can't answer that, the model probably belongs to someone else.
• And your data is yours. If detection requires uploading each build to someone else's cloud, that's a business decision, not a technical detail. Our software is compiled on-machine and works offline.

Six questions for a manufacturer who claims their machine uses AI

This list is just as valid for us as it is for anyone else. Take it to the fair.

1. What exactly does it detect, and with which sensor? “AI monitoring” isn’t the answer. “A 2D build chamber above the powder bed after each coating” is.
2. Does it detect or correct? And if it corrects: within the layer or between layers?
3. What data was the model trained on? Real defects generated intentionally, or simulated ones?
4. Where is it processed? Does any data from my parts leave my network?
5. Can I audit a build? View the layer-by-layer record six months later, when the client asks.
6. Does it replace any testing I have to do now? The honest answer is almost always no, and it’s worth hearing this before buying.

Where are we

Our machines measure and record today: powder bed camera, redundant sensors, and classification of the four faults. Automatic correction is under development, and when it arrives, it will be layer by layer and will not replace inspection. We explain it this way because in this sector, it's often announced as solved with great enthusiasm, and because when we have it, we want people to believe us.

This is about AI inside the machine. The other half—AI before the machine, the one that decides the part's shape—changes something much bigger: when the designer optimizes based on loads, it returns geometries that can't be machined or molded, and additive manufacturing ceases to be the expensive alternative and becomes the only one. It's in when AI designs the parts, it won't design them like we do.

The four steps of process control and where each one is located are in closed loop. What's inside the software and why it's ours is in SamyStudio. And if you're comparing machines, one laser or several is on the other side of the same debate.