Getting your Trinity Audio player ready...
|
Despite the amazing technology advancements that permeate our lives today, there hasn’t been much of a change in how humans deal with the metals for thousands of years. This holds true for wires that move electrical energy in everything from motors to undersea cables, as well as metal rods, tubes, and cubes that give automobiles and trucks their shape, strength, and fuel efficiency. But things are quickly evolving: To enhance current items and develop new ones, the materials production sector is utilizing cutting-edge technologies, procedures, and techniques. In this field of advanced manufacturing, Pacific Northwest National Laboratory (PNNL) is a pioneer.
Scientists at PNNL’s Mathematics for Artificial Reasoning in Science program, for instance, are pioneering cutting-edge approaches to developing and teaching artificially intelligent software to guide the development of cutting-edge production methods. This field of artificial intelligence is known as machine learning. These computer programs are trained to identify patterns in manufacturing data and use this ability to predict or recommend settings in manufacturing processes that will result in materials with better properties than materials made using conventional techniques, such as materials that are lighter, stronger, or more conductive.
According to Keerti Kappagantula, a materials scientist at PNNL, “the components we produce utilizing advanced manufacturing processes are so desirable to business that they want to see these technologies commercialized as rapidly as possible.” Industry partners’ reluctance to invest in new technologies prior to the full development and validation of the underlying physics and other difficulties of advanced manufacturing techniques presents a problem. To close the gap, Kappagantula collaborated with PNNL data scientists Tegan Emerson and Henry Kvinge to create machine learning algorithms that forecast how various manufacturing process variables will affect material attributes. Additionally, the tools convey the forecasts visually so that business partners and others can grasp them right away.
The team thinks it can reduce the time from the lab to the manufacturing floor using these machine learning capabilities from years to months. To find, for instance, what settings result in desired qualities in an aluminum tube, the materials scientists simply need to do a small number of trials, as opposed to dozens. The objective, according to Kvinge, was to employ machine learning as a tool to support the operator of an advanced manufacturing process as they experiment with various device settings and process parameters to find the one that will allow them to accomplish their intended goals.
Solving the right problem
In conventional manufacturing, computer simulations based on a manufacturing process’s well-understood physics allow scientists to see how various parameters affect a material’s qualities. According to Kappagantula, the physics are less known in advanced production. “Deployment is delayed without that understanding.” The Artificial Intelligence Tools for Advanced Manufacturing project by Kappagantula, Kvinge, and Emerson looks for ways that machine learning can be used to extract patterns between process parameters and the resulting material properties. This will give insight into the underlying physics of advanced manufacturing techniques and speed up their adoption.
Understanding how material scientists conceptualize their field—What mental models do they have?—and using that as a scaffold on which to build our models—is the approach that we’ve taken, according to Kvinge. He said that far too frequently, data scientists create solutions to problems that they believe need to be solved rather than problems that other scientists are hoping to solve. Kvinge stated that he believed the team would prefer a machine learning model for this project that anticipated the qualities of a material generated when given particular parameters. He quickly discovered that the materials scientists wanted to be able to provide a property and have a model suggest every possible process parameter that could be utilized to attain it.
An interpretable solution
A machine learning framework was what Kappagantula and her colleagues needed in order to make decisions regarding the next experiment to conduct. Without such direction, developing a material with the appropriate qualities requires trial and error when adjusting the parameters. In order to distinguish between two sets of process parameters and determine which, if either, will produce a material with the desired properties, Kvinge and his colleagues first created a machine learning model called differential property classification. This model makes use of pattern matching capabilities in machine learning to identify which, if either, will produce a material with the desired properties.
Before setting up an experiment, which can be expensive and take a lot of preparation work, the model enables materials scientists to zero in on the ideal settings. Kappagantula advised trusting the model’s recommendation before proceeding with an experiment that the model had suggested. She stated, “I want to be able to observe how it’s performing its analysis. For experts in various fields, the term “interpretability” or “explainability” as it is used in the field of machine learning has diverse connotations. According to Kvinge, the explanation for data scientists and the explanation that makes sense to materials scientists may be completely different for how a machine learning model arrived at its forecast.
Kvinge, Emerson, and their associates approached this issue by attempting to comprehend it from the perspective of a materials scientist. According to Kvinge, “it turns out that they really do comprehend it through these representations of material microstructures.” If you ask them what went wrong, why the experiment failed, or why it succeeded, they will look at the photographs and point out items to you, saying things like, “These grain sizes are too huge, or too small, or whatever.” To ensure their model’s results are understandable, Kvinge, Emerson, and coworkers used images and related data of microstructures from previous experiments to train a machine learning model that generates images of the microstructures that’d result from production process tuned with a given set of parameters. The team is currently validating this model with the goal of incorporating it into a software framework that materials scientists can use to choose which experiments to conduct while creating advanced manufacturing techniques that have the potential to revolutionize the production and properties of materials. Advanced manufacturing, according to Kappagantula, “is not simply producing things more energy efficiently; it’s unlocking qualities and performance that we’ve never seen before.”