PNNL

Picture a drill spinning so fast that the friction it generates can soften and mix metal, allowing it to transform the microstructure of those metals to repair or strengthen them. This advanced manufacturing technology—called friction stir—produces materials with remarkable properties while lowering manufacturing costs.

But friction stir techniques are relatively new, and producing the optimal material can require tricky testing or access to sensors that aren’t feasible on real-world factory floors.

Now, researchers at Pacific Northwest National Laboratory (PNNL) have developed an interpretable, lightweight AI model that can easily predict weld microstructure features using only basic machine sensor inputs, making it easier for manufacturers to get the best results from their friction stir technologies.

Testing the limits of limited testing

“One of the biggest problems when we create high-performance materials using advanced manufacturing approaches is that, often, to evaluate their performance, we have to do destructive testing—and we don’t like destroying samples,” said Keerti Kappagantula, an advisor at PNNL and corresponding author of the paper.

There is another way to evaluate the materials: nondestructive testing using ultrasonic waves; essentially, an ultrasound exam of the material. But there are a couple of problems with that, too.

“We can’t really pause the process to perform ultrasonic testing in a real manufacturing environment,” Kappagantula said. “Ideally, you might have an ultrasonic component for rapid in-line testing, but that’s rare—and even if you have that in-line testing, the heat and vibrations from the friction stir process make it challenging to interpret the in-line ultrasonic signals.”

Instead, Kappagantula explained, friction stir machine operators are typically limited to in-line sensors that detect process temperatures and applied forces—variables that can be controlled using the rotation rate of the friction stir tool and the speed at which it moves through the material.

So, the PNNL team decided to work with that.

A multilayered AI approach

Using a large set of friction stir data, the researchers built a two-step machine learning model. 

In the first step, the model takes the real-time friction stir process parameters—rotation rate and tool traversal—and predicts what in-line ultrasonic testing would show.

Then, based on those predicted ultrasonic data, the model accurately predicts the microstructural features of the finished product (specifically, grain size), which are strong indicators of its mechanical performance.

“The model eliminates the need to perform destructive evaluation,” Kappagantula said. “That means that someone making manufacturing decisions in the real world will not need to destroy samples to assess the quality of the material they’re manufacturing.”

In the future, Kappagantula said, the model could be used to enable automated process control, increasing manufacturing efficiency even further.

Interpretable, lightweight, and adaptable

Throughout the model’s development, interpretability and reliability were priorities.

“We didn’t use a black box model,” said Luke Durell, a data scientist at PNNL and lead author for the paper. “We created a hybrid machine learning model that gave us an understanding of why the analysis showed what it showed, allowing us to see the parameters it was tracking every step of the way.”

“When we work with engineers who are in high-stakes industries—like the nuclear industry—these kinds of models make it easier for them to believe the results and extend it across ecosystems. When you don’t have trust metrics, it’s very hard to get factory floors to adopt AI models,” added Donald Todd, a nuclear engineer at PNNL and a coauthor of the paper.

It was also important to ensure the model could be run on a real-world factory floor. “It’s also a very lightweight model,” Kappagantula said. 

“You don’t need a high-performance computing ecosystem to deploy this,” Durell said.

While the current model is specific to the design, features, and material of the friction stir tool—as well as the characteristics of the sample being manufactured—the researchers say the AI framework could easily be adapted to a wide variety of materials and advanced manufacturing approaches.

“When we use AI for advanced manufacturing, we look to identify an optimal range of process conditions, ensure that we can reliably and confidently identify that range, and enable manufacturers to easily do the same,” Kappagantula said. “This AI model framework meets all of those criteria.”

The funding for this research was provided by PNNL’s Laboratory Directed Research and Development program.