Ready for more disruption in a rapidly changing manufacturing industry? Why artificial intelligence and machine learning are the next big thing in 3D printing
Artificial intelligence (A.I.) has the potential to greatly enhance the capabilities of 3D printing. Possible benefits include improved design optimization, more efficient material usage, faster and more accurate quality control, and the ability to perform predictive maintenance. Overall, the integration of A.I. and 3D printing is expected to lead to more efficient and cost-effective manufacturing processes, as well as new innovations in materials, design, and applications.
The preceding paragraph came from ChatGPT, an online, A.I.-powered service that responded intelligently and appropriately when asked, “How will A.I. affect 3D printing?”
The bot’s answer was long-winded, thus needed some trimming, but A.I. will likely put myself and countless others out of work at some point, as will other forms of automation. For the time being, though, I’m confident in saying that I’m a better writer than a language-model A.I. engine that has but a few months of experience under its digital belt. And with that in mind, I promise that the rest of the words in this article are all human-generated.
Evolving writing skills aside, A.I. and machine learning (ML) shine at tedious, repetitive tasks that would quickly have a human worker yearning for the weekend. One of these is sorting through immense datasets. “Imagine the ability to analyze a huge spreadsheet filled with holes and noise, and then turn that information into a robust model that people can use for decision making. That’s what our software does.”
So said Stephen Warde, responsible for product marketing at Intellegens Ltd., “a spin-off” of England’s University of Cambridge physics department. Today, Warde’s colleagues spend their days developing advanced ML analytics platforms that will soon make additive manufacturing (AM) and other data-hungry technologies smarter, faster, and much less dependent on time-consuming trial-and-improvement experiments and prototyping.
ML and its big brother, A.I., are nothing new. In fact, the first of these terms dates back to 1959, when IBM employee Arthur Samuel trained computers the size of two-car garages to play checkers. Yet ML has long suffered from a severe shortcoming—filling in the gaps—that has only recently been addressed through the work of Intellegens and others.
“Historically, machine learning engines don’t do very well if the data used to train them is incomplete,” said Warde. “Many of us here have a background in physics, materials science, and chemical formulations, and we recognize that sparse data can be a limiting factor. That’s why our team was motivated to build a machine learning method that is successful at building usable models from less-than-optimal data sets.”
That work brought Intellegens to the additive space, where the company’s Alchemite engine can be used for parameter optimization. “Manufacturers typically collect large amounts of data on their processes, the powders they use, laser settings, and so on. Machine learning helps them analyze all this information so they can understand what’s going on in the build chamber and thereby tailor the inputs for the desired results.”
Warde admits the use of this technology overall is still in its infancy, but said he is seeing great interest from the commercial sector, and that additive manufacturers are already using ML software to reduce the need for costly experiments when developing new materials and processes.
Benoit Soete, business development manager for Oqton Inc., a software provider with U.S. offices in Valencia, Calif. and Cary, N.C., is equally optimistic about A.I.’s growing role in AM. “I focus on dental applications here, and can tell you that A.I. use is becoming quite common in dental labs and production centers,” he said.
Crowns and bridges, partial denture frames, aligner models and nightguards—these are some of the typical dental components produced on metal and polymer 3D printers. In terms of mass customization, the dental industry is a glowing example of where AM is headed, but processing all of these unique builds is both time consuming and cumbersome.
According to Soete, skilled technicians must “click a million times” to orient parts, add supports and labels, nest, and otherwise spend countless hours preparing the massive number of dental products built each day. Thankfully, A.I. already eliminates much of this tedium and allows dental technicians to perform more value-added activities.
“Oqton makes the data preparation process highly automated in this and other market sectors,” Soete said. “As A.I. continues to improve, it will assume more responsibility, calculating build parameters based on the part classification, for example, and making suggestions that a human can then accept or override. They might need to make a few tweaks every once in a while, but ultimately, they are getting to a printable file much faster than was previously possible.”
This last statement is crucial. As suggested in the ChatGPT writing example, A.I. still has a long road to travel before it can compete with human intelligence. That, and as Soete noted, “Everyone has their own way of doing things, so if you want to automate, you need a system that is flexible and can adapt to its owner’s reality.”
It should also have the ability to continue learning. Unlike humans, who often forget what they did last year and sometimes forget what they had for breakfast, A.I. can look back at its entire history of decision making and use it to make more informed choices going forward. And when a human overrode those past decisions by reorienting a part or changing a support structure, the A.I. brain will remember this and “get with the program,” so to speak.
“This capability also helps when something new comes along that the A.I. engine might not have been trained on, because it can pretty quickly pick up on what the customer has asked for previously and adjust its logic accordingly,” he added.
Materialise NV of Leuven, Belgium, drinks its own A.I. Kool-Aid. A company known for its long history in AM, Materialise is part service bureau/part software developer, and therefore has an opportunity to evaluate its industry solutions firsthand. CTO Bart Van der Schueren noted that one of these is dentistry, and reiterated many of the A.I. and ML use cases just mentioned.
Automation of the build preparation stage—dental or otherwise—is clearly an essential application for A.I. But given that AM is unique in its ability to create parts one layer at a time, it’s perhaps more important to unravel the mysteries of that process. Doing so would allow manufacturers to build better parts, increase throughput, and eliminate gnawing doubts that can only be resolved through expensive, time-consuming CT scanning and other non-destructive inspection technologies.
One way to do this is to equip 3D printers with visual and thermal cameras, gather images during the build, and then use A.I. to evaluate events such as voids, spatter, cracking, and incomplete melting that might lead to deformation or defective parts.
“Let’s say the recoater blade drags a stripe of powder across a layer halfway through a print job,” Van der Schueren said. “If it occurs in a region between parts, you probably don’t care. And even if it happens directly over the workpiece, you still might not care, provided it corrects itself within the next couple of layers.
“If not, A.I. might be used to stop the process rather than continuing on, wasting time on a part that will end up in the recycle bin and possibly damaging the recoater blade or other machine components,” he continued. “Regardless, capturing and automatically flagging such anomalies allows us to better understand potential problems.”
Using A.I. to identify defects is already a huge improvement, but it becomes even more valuable if used to categorize errors and then prevent them, either through corrective action on the next workpiece or in-situ process adjustments. The latter might alarm those who don’t yet trust a computer algorithm to make decisions that could affect part quality, yet electrical discharge machining (EDM) controls have been applying “fuzzy logic” to electrode paths and parameters for the past few decades—once mastered on a broader scope, such technology would have profound influences on other manufacturing processes: machining, laser cutting, robotic welding, and 3D printing alike.
Here again, A.I. is not yet ready for prime time, although efforts like these are necessary steps in that journey. That’s because they serve ML with what it craves most: data.
“That’s one of the biggest challenges for us and others trying to develop this technology, in that there’s not enough information available to train the models. Even worse, many companies are sitting on their data for obvious confidentiality reasons, so at Materialise we are looking into ways that these companies can learn from their own data without sharing them,” said Van der Schueren.
Despite this, 3D-printer owners and operators can look forward to helpful advice and suggestions from their machines and software systems someday as each grows progressively smarter. That’s great news, but as James Page, vice president of software at Eden Prairie, Minn.-based Stratasys Ltd., noted, “Even if you’ve controlled the process perfectly, you can still end up with a part that doesn’t meet your specs.”
Page is the founder of Riven, a Berkeley-based software developer that Stratasys acquired in 2022. He pointed out that there are many kinds of AM and many opportunities to close the loop, and all are complex in their own way. From powder-bed compaction and thermal cycling of the material to support placement and post-build removal, these and many other factors can make a finished part deviate from its design intent, often significantly, and no amount of A.I. will prevent that. Right?
Not necessarily. “One of the things we’ve been working on is a system that automatically corrects dimensional and geometric errors, whether they come from warping, differential scaling, or even post processing,” said Ward. “We have algorithms now that will correct those and other causes of part variability.”
There’s a caveat, however: The corrective action requires a sacrificial lamb. “We close the loop by printing a test part, scanning it, and then using A.I. to determine whatever processing adjustments are needed to bring the part into specification and keep it there. That way, when you go to make a thousand parts, they all come out with much higher accuracy.”
Ward takes issue with the sacrificial lamb characterization. He said it’s quite common—especially as AM moves into end-use production—to print multiple iterations of a workpiece, gradually dialing in the build parameters and support structures until the process is stable. Using software that can reliably automate this task is a big win. Moreover, the day will come that A.I. will become smart enough to eliminate the test part as well.
“There’s been a huge shift over recent years where 3D printing has transitioned from its traditional role as a prototyping and tooling technology to one of volume manufacturing,” he said. “But one of the challenges that the manufacturing industry in general continues to face is AM’s inability to meet stringent dimensional tolerances. With the predictive models and algorithms we’re launching, we’re going to change that and open up entirely new classes of parts that additive couldn’t previously approach.”
All this talk of what’s coming down the AM pike is exciting stuff, but as Trent Still likes to point out, many of those reading this already use A.I. every day, even if they don’t refer to it as such. Still, the senior manager of technical marketing for design and manufacturing at San Francisco-based Autodesk Inc., is talking about generative design software, a tool that goes hand in hand with AM and its ability to build pretty much whatever is thrown at it. As with A.I., this nascent technology is poised to get a whole lot smarter.
“The next step is what we call generative modeling, which is effectively a manufacturing-aware tool that allows industrial designers to arrive at optimal solutions within seconds,” he said. “Historically, generative design required you to be an experienced engineer or simulations expert to set up a modeling study, but we’ve created new A.I.-powered engines that can learn over time how to satisfy various design needs.”
One example of this comes from an aftermarket automotive parts supplier that needed to manufacture an exhaust manifold connecting several components inside an “exceptionally tight” engine compartment. “Previously, it might have taken them a month to model, 3D print, and test a handful of prototypes, but A.I. produced three structurally viable designs within a couple of hours.”
These are impressive capabilities, but they’re nowhere near A.I. and ML’s north star, which is to have them consume various structural, operational, and electromechanical requirements and not only deliver an optimal product design, but also tell the engineer how and where to manufacture it.
Taken to the extreme, A.I. might even design production floors. “Let’s say you’re a Tier One manufacturer planning a multi-billion-dollar facility,” Still said. “Such an undertaking means answering thousands or perhaps millions of very specific questions so as to arrive at an optimized throughput for your factory. A.I. can help manufacturers make more informed decisions for problem number 1,267, even though they’re only on problem 37 right now. It will also allow us to unify historically isolated industries like architecture, engineering, and construction, improving their ability to share information and insights.”
Sakthivel Arumugam is a lead product manager at Ansys Inc.’s Cambridge, U.K., location. The Canonsburg, Pa.-based software provider focuses on engineering simulation, and as indicated several times so far, A.I. is becoming an integral part of many simulation software tools, whether it’s simulating the results of decisions made during product design, simulating the manufacturing process, or simulating the behavior of materials, parts, and assemblies under various operating conditions.
In Ansys’ case, these capabilities extend beyond pure simulation to include the capture, analysis, and management of engineering data, including that associated with the raw materials used in AM and other processes. “Materials are a huge factor when it comes to the qualification and certification of additively manufacturing parts, especially in the aerospace and defense sector,” Arumugam said. “To support this market, Ansys has a business unit devoted entirely to a materials data management solution called Granta MI.”
AM is a very data-heavy process, he added. For starters, there are the raw material properties and characteristics to consider, whether it’s metal or polymer, and whether it comes to the factory floor in resin, powder, wire, rod, sheet, bar, or filament form.
But just as important is what comes after. These characteristics change as the material melts, cures, or fuses, and quite often, change once again when the part moves on to secondary heat treatment or curing. Further complicating an already complex situation is the fact that the product’s dimensional and mechanical properties can be further influenced by the strategy used to build the part, whether it’s laser power, cross-hatching approach, and so on. It’s Granta MI’s job to capture and manage all this data, explained Arumugam.
“Customers can gather data from each step of the printing process and enter it into the platform’s database in a very organized, integrated manner. And since it’s all linked, it then becomes easier for the user to understand the relationship between a finished material property, say, and the process parameters, what machine was used, and even the material supplier.”
Where do ML and A.I. fit into this equation? Analyzing data like this is one of the boring, repetitive tasks that A.I. excels at. As such, Ansys partnered with Intellegens and incorporated its algorithms into the Granta MI platform.
“This is yet another extension of materials data analytics,” said Arumugam.
“Manufacturers can now gather, analyze, and store data about the raw materials and the relationships that exist between their properties and the final product. They can use this information to produce higher quality parts or to optimize their processes, and if there’s a product failure years down the road, a forensic analysis can then be performed to understand why, with the resulting data being used for continuous improvement purposes. It’s a big job and a lot of data, but A.I. helps make it all manageable.”
Connect With Us