AI has slowly been incorporated into the 3D print universe, typically at rather low levels: a 3D printer might be able to detect spaghetti printing, for example, or a slicer might adjust settings to make a part stronger.
While those developments can make life a bit easier for the 3D printer operator, we still need the 3D printer operator. That role coordinates all the myriad of steps required to get from A to B in a 3D print workflow.
The goal is to produce a good part in the end, but there are multiple steps along the way: slicing, job setup, monitoring, etc. Typically, 3D printer operators will iterate printing several times to determine the optimal parameters for the specific part. That’s labour-intensive, and therefore prone to error and requires skilled operators.
I’m reading a research paper from CMU that proposes a method to have an AI, an LLM to be specific, to completely run through the typical FFF workflow.
They built a multi-agent LLM supervisor that watches FFF prints with two cameras (top + front). After each layer, the supervisor decides if there’s a defect, queries the printer (via API), and then autonomously tweaks parameters (flow, speed, retraction, Z-offset, temps, etc.) before resuming. There is no fine-tuning; it relies on in-context prompting.
The experiments were performed on a stock Creality Ender-5 desktop FFF 3D printer, and used PLA and TPU. One modification to the print jobs was to include a prime tower that would handle the ooze generated while layers were being inspected.
What happened? The system detected stringing, under/over-extrusion, poor adhesion/layer separation, and warping. It suggested and executed changes like a slower print speed (~75%), increased material flow >100%, pressure advance tweaks (~0.1), retraction tuning (~25% flow / faster retract), and higher TPU nozzle temps (~220C).
These are parameter tweaks that 3D printer operators would very typically do as they observe the initial prints of an object with a new material. However, here instead of the parameters being changed manually, they are changed by the LLM automatically.
This is an impressive result, and one that I hope makes its way into commercial equipment. Today we have 3D printers that have closed-loop controls, but this technology takes that to an entirely different level. One can imagine future 3D printers that are able to print perfect parts, even if you are using new materials that haven’t been tuned.
Via ArXiv
