No More 3D Print Failures Thanks To... AI?

In this episode, we explore how researchers have developed an AI system that can watch a 3D print as it’s being made, detect mistakes, and automatically correct them before the print fails. Learn how this breakthrough could dramatically reduce wasted material, improve print quality, and move manufacturing one step closer to fully autonomous production.

Farbod: Hello friends and welcome back to the Next Byte podcast. Let me ask you a question that might be important to you, but is definitely important to me. What if your 3D printer could catch mistakes before the print fails? Well, today we're diving into a breakthrough where AI powered systems monitor the prints in real time, spotting defects layer by layer and automatically correcting them before wasted plastic, hours and failed builds pile up. It's a glimpse into the future of smarter autonomous manufacturing and a huge step away from the failed prints that I grew up with. Let's get into it.

Welcome back friends. As you heard today, we're talking about additive manufacturing. But before we get into that, let's just talk about manufacturing in general. We have this resource that's talking about industrial automation, some of the best practices to keep in mind, and it's from one of our favorite sponsors, and that's Mouser Electronics. Mouser is one of the world's biggest electronic distributors, and one of the things that we like about them is the look at what's trending, what's high impact, and they share insights from either academia or industry that are relevant to the conversations we're having. So this one is super timely. Everyone's kind of talking about industrial automation. Manufacturing's a hot topic. And it feels like every day I'm seeing something about robotic arms being deployed or the next, I don't know, factory 4.0 technology coming out and something to keep in mind, the number one principle is always safety.  How we employ machine safety and how do we make sure that anyone who's working there is staying safe. This resource kind of goes over that from A to Z. They talk about how to prevent hazards, prevent incidents, how to leverage technologies like e-stops, mats, and sensors to kind of create this close feedback loop for safety. And I don't know, I think it's just a good asset to have, a good resource to have as you're going into these conversations or doing a gut check on any tweet, any article, any podcast you might be listening to that's about this topic.

Daniel: It's a good principle for machine design in general. is having sensors, having electronics, having a brain to help close the loop on what's actually going on and make sure that there aren't safety incidents. It's an important principle in the same way that people need closed feedback loops for quality, which is something we're going to talk about during today's episode. Obviously, Mouser hits the nail on the head again with a very useful, practical, technical resource that we're going to link in the show notes and you should check it out.

Farbod: As the kids say, they don't miss. With that non-miss already loaded up at the beginning of the episode, let's venture onto Carnegie Mellon University. I feel like we haven't talked about them at all in the last year or so, and that's unfortunate because Carnegie Mellon has had some of my favorite topics that we talked about on the podcast, especially for 3D printing. So we're going back there for the same reason. This research is coming from PhD candidate Yayati Jathav and their professor, Professor Farimani.  They're doing self-correcting 3D printing or additive manufacturing. And if you've ever touched a 3D printer before in your life, that already like resonates with you, just like the headline of it, right? So for those of you that don't know, every time you use a 3D printer, you have your design and everything ready. You put it into this software, which is called a slicer, and that generates some sort of machine code that tells the printer what to do. Within that machine code, you're setting parameters like the temperature of the nozzle that's depositing, for example, the molten plastic, or the temperature of the bed or the flow rates, et cetera, et cetera, all of which impact your print. Now the tricky bit is that as your materials change, as you go from one printer to another, these variables have to be adjusted so that you can actually have a stable print that doesn't fail. People like Daniel and I, which kind of have been using this manual trial and error process, know all too well that it's very, very tedious to get your settings right at least with today's technology.

Daniel: Well, the best practice is like, you you've got all these knobs you can turn, nozzle temperature, print speed, layer height, extrusion rate, cooling fans, stuff like that. the, the like best practice or like the advice you get from people on the internet, if you have failed prints, it's like, you know, just try a test print, you know, try this, try that, see if it works. If it does, great, scale it up. It might work for your big print. It might not and by the way, if the environmental conditions change or the materials change or the age of the material with which you're printing changes, all these things might need to change too. It's basically a black box and it sucks because you know, you could have a print that works great two out of every five times and you can run with the same settings again and it runs like crap two of the other five times and then the other ones hit or miss. And it's really challenging to like put your finger on the mechanism behind all these different things because they're all so interrelated. And there's not a great way of testing or adjusting things in real time while you're printing to make sure that if an issue starts to happen that you correct it. Typically, if you have a bad part, it's just a bad part and you have to throw that away and start again from scratch.

Farbod: And the reason that that's the case is because it's an open feedback loop system. What we mean by that is that the user is inputting something, then the controller's starting a program, and then it runs. There is no sort of intervention that's happening in there that's constantly giving feedback of like, this went well or this didn't go well. So with that in mind, the hope that this research had was can we introduce some sort of a closed feedback loop system which would prevent us having these failed prints and kind of getting it right in the first pass. Now, were they able to do it? How do they do it? This is our favorite part. As you might be able to tell either by the title or just, you know, the way that the world has been for the last five years, there's gonna be an AI element here. Everything has an AI element these days. And even the researchers said it themselves. This kind of technology would have been really difficult to implement had it not been for some of the advances we've made in AI, and specifically large language models in the past couple years. So, with that in mind, here's the strategy that they had. Obviously we need some sort of a visual inspection, as it exists right now. If Daniel and I were doing a print, we would know if it's going well or not by just looking at it. So the idea was, let's place a camera, let's look at the front, let's look at the top of the print, and every time we print a layer, we can kind of do an analysis of if things are going well. Then that data is going to a multi-agent framework. Here you have ⁓ three core agents that are specialist experts and then one supervising agent. And the way they describe it is the supervisor agent is kind of like the maestro at an orchestra that calls on every segment to do a specific task. So going through it from the top, you have this visual data coming in and then you have a vision language model that can ingest that and in human terms, so using common language, say what these errors are. This can then go to the supervisor and the supervisor can pass it to the planner. The planner can say, based on that information, I think we should be adjusting the flow rate, the fan speed, the temperatures, et cetera, et cetera, which then goes back to the supervisor. And then the supervisor can tell the executor agents, hey, I need you to convert all these recommendations that are given again in human language, human terms and convert it to machine code, right? And then the executor does that and shoves it to the printer, which now has a new set of parameters to print with and do this all over again. So that's one round of what this process looks like to give you kind of like an overview.

Daniel: I think this is interesting because a lot of the systems on the 3D printers already, well, two pieces of feedback here. One of them being a lot of the systems on the printers already will use a PID controller to ensure that that specific subcomponent is performing the way it's supposed to. So when you set a heating nozzle temperature, a PID controller helps adjust the amount of power taken to that heater to make sure that the nozzle is at the correct set temperature. does the same thing for fan speed. It does the same thing for bed temperature and even for the positions of the motors. But those are all just controlling and doing a closed feedback loop system within that subsystem, not the entire system of the whole. It's like, how is the quality of the output of this entire system supposed to be, which is like the geometry of the part. How can we fix that? That's awesome.

The second piece of feedback, it's kind of like a little bit of like, devil's advocate criticism here, is I did look and there's a lot of other 3D printers that already have some sort of like built-in quality detection on the part during production. So really, really high-end additive manufacturing machines use things like laser monitoring. They have cameras. They'll also put like a layer of thermal cameras on top and do inspection on every single layer and do like metrology basically while the parts being produced. But the problem is most of these still only do detection. They don't do autonomous correction. So what's really cool about this, like LLM driven agentic reasoning layer that this team built is I think this could even potentially be retrofitted onto printers where the hardware is already there to do the sensing. And you just add an additional reasoning layer that can potentially correct failed prints mid-print, as opposed to doing what they do now, which is stop the print and allow you to start from scratch. I think this is like a really interesting application, not just for retrofitting sensors and hardware to existing printers that don't have these closed loop systems, but then also for the ones that do have closed loop systems using it as a part to, not just being interrupt to the system, but also going back into the control and potentially self-correcting the print halfway through.

Farbod: You hit the nail on the head there in terms of the comparison with existing technologies. But another thing I'm gonna add is one, this mentality of open sourcing it so that everyone can keep improving on the technology. But two, one of the things I appreciated is that in this multi-agent framework, they specifically architected it so it was palatable to manufacturers, right? Like you're saying, every one of these manufacturers has their own proprietary way of understanding if the print is going bad. They also understood that if this was gonna be adopted across the board, those proprietary pieces should have still remained there, right? So imagine the piece that's talking to the actual machine, interpreting the human language to the machine code. Maybe whatever manufacturer there is out there does not wanna expose that to the community.

They architected it in such a way that the rest of it can plug and play with whatever the manufacturer provides, almost like a driver. Which means that everyone can kind of keep tweaking this model, making sure it's working well, and manufacturers can reap all the benefits of it, of having this standardized layer. So whether it's an older machine that has no support or a new machine that's coming out and would like to benefit from something like this as an added layer, everyone can reap the benefits without giving up any of the proprietary tech.

Daniel: When I like the names of the agents like planner, executor, supervisor, those are job titles that exist in an actual print farm. If someone is running 3d printers to produce parts. And it's true that those people don't actually need to understand what's going on proprietary inside the brain of these machines to make sure that they're running well. This is just using technology with cameras to help inform that instead of the planner executor supervisor loop, checking quality at the end of the production of every single part, which is what happens right now. They're able to do the plan, execute, supervise, mid print. So in between layers to make sure that you can correct a part if it starts to drift as opposed to waiting for that part to be complete before you realize that it's

Farbod: Exactly. Just kind of some important so what's out of this. The parts that they printed using this technology showed a 5x increase in peak load capacity. So much better structural integrity than one printed from the best practices that you and I were talking about. Given material, given technology, given slicer, use these parameters, it already beat it out. And then this one I thought was even more interesting. This model consistently identified failure points when compared to against other additive manufacturing experts. They tested against 14 other people, I think. So not like you or me, but people that do this for a living, for example. In terms of how they would set up a print, in terms of how they would detect defects, the machine, sorry, the model did a really good job of consistently picking up those failure points as well. And then the last bit I think is interesting for like a tangential benefit of what's going on with these machines where every time the model is making a correction, it's actually outputting commentary on what it thinks is wrong and the updates that it's making. So whether you're a manufacturer of those machines or just an engineer that's tweaking around with it, you now have some context in terms of what's going on and the corrective measures taken instead of it being this black box of it just does stuff for me and I look away from.

Daniel: And that last part I think is the most important part to me when it comes to like me putting myself in the shoes of someone who owns, you know, a several hundred thousand dollar 3D printer and understanding whether I'm going to actually use this or not. It's like having live feedback from the model that gives you an idea of what's going wrong and what the corrections are that it's putting into place helps me personally reduce my skepticism toward the amount of like the extent to which hallucination of these AI agents might play a role in the corrections that it's making that I think that's a risk with using any type of large language model, which tends to hallucinate. It helps me when I'm working with a development tool like cloud code as example, to watch its reasoning and to watch the changes that it's making. Even if I'm not the one directly typing into the code base, helps me to watch that and ensure that what I'm seeing and the reasoning that is associated with that checks out to my intuitive understanding of what's going on. I think there's a lot of 3D printing experts that could potentially find opportunities to improve the model or at least give themselves some reassurance that the model is doing the right thing if they're able to watch this commentary while it's happening live and use that as a way of understanding how and why this AI brain is working the way it is.

Farbod: Yep, totally agree with you. Kind of to wrap up everything that we talked about, ⁓ we have these researchers from Carnegie Mellon University who built a LLM powered system that fixes 3D print errors that are happening in real time. The reason they did this is because most 3D printing systems are open looped. So you can't really make changes automatically unless you're physically being involved modifying, iterating, manual trial and error basically.

So they did this by having cameras that capture each layer from the top and from the front of the print during printing, then a vision language model that detects defects and assesses the print quality. And then they have this four specialized LLM agents that plan, execute, and orchestrate the corrections that need to go into the printer. And what it results in is basically parts with much better structural integrity. We're talking about a five-time increase in peak load capacity.

And the reason that this is important is that it could reduce print failures and automate quality so that you never have to worry about the trial and error that Daniel and I did 10 years ago.

This episode was brought to you by Mouser, our go-to source for electronics parts for any hobby or prototype. Click HERE to learn the machine safety best practices that keep modern industrial automation running safely, reliably, and efficiently.

Become a founding reader of our newsletter: http://read.thenextbyte.com/

As always, you can find these and other interesting & impactful engineering articles on Wevolver.com.