Machine learning accelerates advanced manufacturing technologies

​Although life today is filled with astonishing technological advances, the use of metals that underpin these developments has not changed significantly in thousands of years. That's everything from the metal rods, tubes and cubes that give cars and trucks their shape, strength and fuel economy, to the wires that carry electricity to everything from power plants to undersea cables.

But things are changing rapidly: the materials manufacturing industry is using new innovative technologies, processes and methods to improve existing products and create new ones. The United States' Pacific Northwest National Laboratory (PNNL) is a leader in this field, known as advanced manufacturing. Founded in 1965, PNNL leverages its unique strengths in chemistry, earth sciences, biology and data science to advance scientific knowledge to address sustainable energy and national security challenges.

Scientists working in PNNL’s Artificial Intelligence Reasoning in Science project have pioneered ways to design and train computer software using machine learning, a branch of artificial intelligence, to guide the development of new manufacturing processes.

These software programs are trained to recognize patterns in manufacturing data and use this pattern recognition ability to recommend or predict settings in the manufacturing process that will produce materials with improved properties than using traditional methods The materials produced are lighter, stronger or more conductive.

Keerti Kappagantula, a materials scientist at PNNL, said: "The components we create using advanced manufacturing processes are very attractive to industrial companies, and they want to see these technologies launched as soon as possible."

A challenge However, industry partners are reluctant to invest in new technologies until the underlying physics and other complexities of advanced manufacturing technologies have been fully fleshed out and proven.

To bridge the gap, Kappagantula worked with PNNL data scientists Henry Kvinge and Tegan Emerson to develop machine learning tools that predict how various settings in the manufacturing process affect material properties. These tools also display forecasts visually, providing immediate clarity and understanding to industry partners and others.

By using these machine learning tools, the team believes the time from lab to factory can be reduced to months instead of years. Guided by the tool's predictions, materials scientists can determine future material properties by conducting only a few experiments, rather than dozens. For example, what settings would cause aluminum pipes to perform as expected.

Our goal is to use machine learning as a tool to help guide people who are running advanced manufacturing processes to try different settings — different process parameters — on their equipment to find One that allows them to achieve what they actually want to achieve."

solve the right problem

In traditional manufacturing, computer models are built on a very good understanding of the physics of the manufacturing process , showing how different settings affect the material's properties. Kappagantula said that in advanced manufacturing, the physics is poorly understood. Without this professional understanding, production is delayed.

New Advanced Manufacturing Artificial Intelligence Tools project aims to identify how machine learning can be used to extract patterns between process parameters and resulting material properties, which provides insights into the underlying physics of advanced manufacturing technologies and could accelerate their deployment.

"The approach we take, the unifying theme, starts with understanding how materials scientists apply their expertise and what mental models do they have? And then using that as a framework to build models," Kvinge said.

In this project a machine learning model is required to predict the performance of a material given specific parameters. In consultations with materials scientists, he quickly learned that what they really wanted was to be able to specify a property and have a model suggest all the process parameters that could be used to achieve that property.

An Explainable Solution

What Kappagantula and her colleagues needed was a machine learning framework that could provide results that would help her team make decisions about what experiments to try next. In the absence of such guidance, the process of adjusting parameters to develop materials with desired properties is fraught with risk of failure.

In this project, Kvinge and his colleagues first developed a machine learning model called "differential attribute classification", which uses the pattern matching capabilities of machine learning to distinguish two sets of process parameters to Determine which group is more likely to produce a material with the desired properties.

The model allows materials scientists to lock in optimal parameters before starting experiments, which can be expensive and require extensive preparation.

Kappagantula said that before conducting experiments on machine learning model recommendations, she needed to trust the model’s recommendations. "I'd like to be able to see how it performs analysis."

This concept is called interpretability in the field of machine learning, and it means different things to experts in different fields. Kvinge noted that for a data scientist, the explanation of how a machine learning model arrived at its predictions may be completely different from the explanation that makes sense to a materials scientist.

When Kvinge, Emerson and their colleagues tackled this problem, they tried to understand it from the perspective of a materials scientist.

"It turns out they know this very well from pictures of the microstructure of these materials," Kvinge said. "If you ask them what went wrong, why the experiment didn't go well, or why it went well, they They'll look at the pictures and point out problems to you, saying these particles are too big, or too small, or something like that." To make the results of their machine learning model interpretable, Kvinge , Emerson and colleagues used microstructural images and related data from previous experiments to train a model that generates microstructural images that would result from a manufacturing process tuned by a given set of parameters.

The team is currently validating the model and working to make it part of a software framework that materials scientists can use to determine which experiments to conduct while developing advanced manufacturing technologies that promise to transform materials production and performance.

Kappagantula said of advanced manufacturing: “It’s not just improving energy efficiency, it’s opening up new materials with never-before-seen properties and performance.”​

The above is the detailed content of Machine learning accelerates advanced manufacturing technologies. For more information, please follow other related articles on the PHP Chinese website!