The additive manufacturing industry embraces AI

Companies in the 3D printing sector are rapidly adopting artificial intelligence technologies to further extend additive manufacturing’s capabilities. Machine learning (ML), a subset of AI in which a computer “learns” to perform a task without being explicitly programmed to do so, is the first area in which businesses are developing practical applications.

“Some 3D printing companies use machine learning to help design parts, and some use it to help identify which parts are a good fit for additive,” said Zach Simkin, president of Senvol, a New York City company that supplies software and other data-related products for implementing AM.

“Part of what’s unique about additive manufacturing is you’re effectively building the material as you’re processing the part, so material properties can vary substantially,” said Simkin. “Changing the parameters, such as laser power, scan speed, or hatch spacing, can have a significant impact on the part’s properties.”

Typically, determining the optimal parameters for generating a part with the desired properties has required expensive and time-consuming trial and error.

Data Versus Physics Models

Senvol’s ML software takes a data-driven approach. That means it works for any type of additive machine, any material, and any process—be it metal alloy on a powder bed fusion printer or a polymer filament on an extrusion machine.

“Our software analyzes the relationships between inputs and outputs, but it is entirely up to the user to define those inputs and outputs,” said Simkin. So, for example, it can analyze how process-parameter data impacts material properties or how feedstock characteristics affect mechanical performance.

“We use empirical data as both input and output, and our software then ‘learns,’ or infers, what those relationships are,” said Simkin. This data-driven approach is juxtaposed against the more common types of modeling, which rely on physics-based formulas to calculate output.

Simkin believes the two approaches are complementary. “Physics-based models can take hours or days for a simulation, depending upon the fidelity of the model, but Senvol ML software runs simulations on a standard computer—sometimes in a matter of seconds,” he said. On the other hand, physics-based models show causation, whereas Senvol’s data-based approach only shows correlation.

“The physics in additive manufacturing today are not as well understood as we would ultimately like, so developing and validating those models can take a long time,” Simkin said, adding that the company’s data-driven approach is “being used today to develop better process parameters with far fewer builds.”

Fixing Repeatable Errors

Another company incorporating AI technology is Inkbit, a Medford, Mass., startup founded on the idea of joining machine vision and AI to make production-ready AM systems.

The multimaterial, inkjet-based AM platform uses ML software in combination with a 3D machine vision system to provide the reliability, throughput, and precision required on a factory floor.

All 3D printing processes experience errors, including repeatable statistical errors such as material shrinkage and rounding of the edges.

“Because of these statistical errors, you don’t get a printed part with exactly the same geometry as the file you submitted,” said Wojciech Matusik, cofounder and CTO of Inkbit and an MIT professor of electrical engineering and computer science.

“But with our system, you send a file to the printer, and then our 3D scanner measures precisely the part that comes out,” said Matusik. “This creates a data set with both the input geometry and the output geometry.”

Comparing the two datasets teaches the machine to how to best process the material. Then, by reversing the data stream, the machine learns with an inverse model to create a predistortion of the input geometry.

“Imagine that your machine prints a file that is supposed to be 10 centimeters in length, but due to material shrinkage, the part comes out at 9.9 cm,” Matusik said. “Our machine-learning-inversion process will figure out that in order to get parts that are 10 cm long, you have to send the printer a part file that is 10.1 cm in size.”

The ML system runs in the back end of the Inkbit system, automatically correcting process parameters to generate a part that meets the initial requirements.

“At Inkbit, we believe that machine learning will become more and more common in 3D printing systems, because the results we’re seeing are pretty incredible,” said Davide Marini, cofounder and CEO at Inkbit. “This system allows you to get a new level of printing accuracy and precision that you cannot achieve without machine learning, so I think every company—at some point in the future—will be forced to adopt some of these approaches.”