In Sir Arthur Conan Doyle’s “A Study in Scarlet,” Sherlock Holmes told us, “It is a capital mistake to theorize before one has data.” It’s a great quote. Here’s another: “You can’t improve what you don’t measure.” That’s according to management consultant Peter Drucker, who also said, “The best way to predict your future is to create it.”
Zach Simkin would agree on all three points. The co-founder and president of New York-based Senvol, he and business partner Annie Wang have spent the past decade developing products and services that help manufacturers access, generate, and analyze additive manufacturing (AM) data.
“I was working on my master’s degree when the class received an assignment on 3D printing,” said Simkin. “That was my first real introduction to additive, and I was enthralled instantly with its possibilities. Coincidentally, that was around the time I met Annie, who had similar interests. We began asking questions and gathering as much knowledge and information as we could, and it soon became evident that there was this huge gap in the amount of data available to additive manufacturers. That’s when the lightbulb went off.”
Senvol was born as a result of that lightbulb. Since then, the company has made a name for itself, not only as a purveyor of 3D printing data and related software tools, but as a partner to the United States Department of Defense (DoD) and its various branches, as well as manufacturers such as Northrop Grumman. Simkin has also been active on the SME AM Advisory Board, serving as vice-chair and then chair, and as a member of ASME’s Technical Advisory Panel (TAP) for Additive Manufacturing.
He attributes some of their early success to Jim Williams, owner of Paramount Industries, a patternmaking shop turned 3D-printing service bureau. 3D Systems acquired Paramount in 2012, and Williams served as the vice-president of aerospace and defense there for a spell, but back when Simkin and Wang were still master’s students, he took them under his wing and opened some doors that otherwise would have remained firmly shut. This included additive manufacturing technology and education provider America Makes, which Simkin said “was very welcoming to us, despite being new themselves at that time. Since then, that partnership with and membership in America Makes has proven very valuable.”
Much of Senvol’s work is aimed at helping companies select or qualify materials for 3D printing and evaluating associated process parameters. Its products include Senvol’s database of additive machines and materials (known as the Senvol Database), an application programming interface (API) that software developers can use to access that database, and various libraries of standard operating procedures. The company also offers AM data sets called Senvol Indexes, and its latest product, Senvol ML, a “data-driven machine learning software to analyze the relationships between additive process parameters and material performance.”
In one recent example, Senvol collaborated with metrology giant ZEISS and AM product and service provider Materialise to assist Rosswag Engineering, an industrial forging company in Germany, with streamlining the qualification process for the feedstocks used on its laser powder bed fusion (LPBF) printers.
Previously, Rosswag technicians had managed to qualify more than 40 materials on their own, but it was time-consuming, very manual work. Senvol provided them with its Senvol ML machine learning software, which served to automate the design of experiments setup and then analyze the results.
Together with the Materialise Process Tuner (MPT) and ZEISS AM parameter workflow tool, Senvol ML helped to “significantly reduce time, costs, and the number of builds needed to determine the optimum print recipe,” according to a company spokesperson.
“Let’s say you want to print 50 test coupons, each with a different parameter set,” said Simkin. “Without an automation tool such as the Materialise Process Tuner, an operator has to manually key in the parameters one at a time for each coupon. This not only takes a long time but is error-prone. And each of those coupons must be evaluated after printing, most often through CT scanning, and the resulting data gathered and interrogated. Between those steps and the initial design of experiments—which can also be quite time consuming—we were collectively able to optimize the entire process.”
Ironically, the collaboration happened by accident—none of the companies knew that Rosswag was using all three solutions independently. It was only after this fortunate discovery that Senvol, Materialise, and ZEISS got together to develop a joint workflow and publish the results.
Note the term “machine learning” used earlier. It’s an important one to the AM community for a number of reasons. Simkin ticked off a few of the process variables present in some AM operations, among them build speed, layer thickness, infill density, and part geometry. These are the inputs, he said, which are linked to process outputs such as the finished part’s surface roughness, tensile strength, flexural modulus, and much more. Evaluating each of these in an attempt to determine what inputs are required to achieve the desired output is a gargantuan task for the human mind. It’s understandable, then, that we’re not very good at it, a factor that is holding the industry back.
As proof, Simkin referenced another project, this one with Northrop Grumman and funded by America Makes, during which he pitted his company’s machine learning software against Northrop Grumman’s team of experienced AM process engineers, who followed a traditional process for parameter optimization. The task? To determine the optimal 3D printing parameters for a specific list of workpiece performance requirements.
Sadly, the man vs. machine argument wasn’t even close. Simkin’s take: “The short of it was that none of the performance requirements for any of the applications we evaluated were met by the parameters that Northrop Grumman had selected, and literally every single objective for every goal was met by the parameters selected by our machine learning software. It was night and day.”
The reason is twofold, he explained. For starters, Northrop Grumman engineers had access to the same data sets used by Senvol’s machine learning software but were unable to analyze that mountain of data in a meaningful manner, making it impossible to build a comprehensive and accurate process model—they couldn’t fill in the gaps, so to speak. And who can blame them? Simkin noted that, in this particular case, roughly 220,000 possible parameter combinations existed, of which Northrop Grumman was only able to evaluate a few hundred—the parameter combinations for which they had empirical data.
The second issue was that Northrop Grumman assumed that each empirical data point represented an average value for how that particular parameter set would perform or what it would deliver. That is not always true, however. “Even if you do repeat analyses of the exact same parameter set, there’s going to be a distribution curve, with some points providing better results than others,” Simkin said. “This is exactly the kind of task that computers are good at and humans aren’t, which is why machine learning will play an increasingly important role in these types of additive manufacturing scenarios.”
One final argument comes from BMW, where AM engineers struggled to find optimal parameter sets using a traditional development approach. After selecting parameters that they had deemed optimal (though unsatisfactory), they engaged with Simkin and his team, who opted to begin the selection process from scratch. The results—to the automaker, at least—were surprising.
“We expected a positive outcome, but it was a fun project nonetheless because it gave us a true apples-to-apples comparison of the parameters obtained with machine learning software vs. those without,” Simkin said. “The former clearly outperformed the latter, and, what’s more, allowed the customer to produce a family of parts they had previously been unsuccessful in building because they had not been able to identify a parameter set that would work.”
The ability to quickly generate and optimize parameter sets and then fine-tune them for the material, part geometry, and 3D printer presents some exciting doors. Greater productivity and part quality sit behind one of them, but behind another is the potential to adjust variables mid-process, catering them to the constantly changing geometry as the part grows and the platform works its way downward during the build.
“In-process adjustment of build parameters represents a very advanced use of the technology that few in the AM community are doing today, at least based on what I see,” said Simkin. “But I do think that, certainly, the industry is heading in that direction. We’ve worked on some programs to explore the possibility further and have even supported a few organizations that are experimenting along those lines, but it’s clearly not something done by the masses. At least, not yet.”
Simkin noted that machine learning does more than speed up material development and parameter optimization. It also expedites the qualification process, for machines and materials alike. For instance, in a recent Air Force research program with the University of Dayton Research Institute (UDRI), Senvol helped qualify a recently purchased M400-4 quad-laser printer from EOS. “We used machine learning as a tool to accomplish qualification more efficiently, accurately, and a whole lot faster.”
Looking back at the past ten years and all that he and his team have accomplished, Simkin feels that we as a community are at an especially interesting yet unique time in the industry. The founders of 3D printing technology are still with us, he said. They’re still walking the show floors, some are going to work each day, and practically all of them are still accessible. “I think it’s really cool that we can reach out to these folks and talk to them and learn from them. A generation from now, people won’t have that opportunity. We must document what these pioneers have done while they’re still with us.”
Someday, and sooner than any of us would like, Simkin and many of those reading this will be among these pioneers. “I’ve only been at this for a decade and some of the newcomers to additive manufacturing consider me an old-timer,” he laughed. “The days sure go by, don’t they?”
