Exploring the inside of materials from the outside with a deep learning system

You may not be able to tell a book by its cover, but MIT researchers say it could be possible to do the same with all sorts of materials, from airplane parts to medical implants. Their new approach allows engineers to understand what’s going on inside a material simply by looking at its surface properties.

Using a form of machine learning known as deep learning, the team compares large amounts of simulated data about the material’s external force field and the corresponding internal structure, and uses it to generate reliable surface-to-interior I have generated a system that can make predictions on the data.

The result was published in a magazine advanced materialsa paper by PhD student Zhenze Yang and Professor Markus Buehler of Civil and Environmental Engineering.

“This is a very common problem in engineering,” explains Buehler. “If you have a material, such as a car door or an airplane part, and you want to know the inside of the material, you can take an image and calculate the amount of deformation to measure the strain on the surface. You can’t really see it, and the only way to know for sure is to cut the material and then look inside to see if there’s any damage there.

X-rays and other techniques could be used, but these tend to require expensive and bulky equipment, he says. The thing is: Can we develop AI algorithms that can look at what’s going on at the surface? Are we looking at the surface of matter and trying to understand what’s really going on inside? may include material damage, cracks, stresses, or internal microstructural details.

The same sort of problem could apply to living tissue, he adds. “Is there a disease there, or is there some growth or change in the tissue?” The aim was to develop a system that could answer these kinds of questions in a completely non-invasive way.

Achieving that goal required dealing with complexities, including the fact that “many such problems have multiple solutions,” says Buehler. For example, many different internal configurations may exhibit the same surface properties. To address that ambiguity, “basically, I created a method that can give you all the possibilities, all the options, that could result in this particular outcome. [surface] scenario. ”

The techniques they developed involved training an AI model using vast amounts of data about surface measurements and related internal properties. There were not only uniform materials, but also combinations of different materials. “Some new planes are made of composites, so they are intentionally designed to have different phases,” Buehler says. “And of course, in biology, all kinds of biological substances are made up of multiple components, and they have very different properties. There is substance.”

The technique works, he says, even for materials whose complexity is not fully understood. “With complex biological tissues, we cannot understand exactly how they behave, but we can measure their behavior. We don’t have a theory, but if enough data is collected, we can train a model.” .”

Yang says the methods they have developed are broadly applicable. “It’s not just limited to solid mechanics problems, it can be applied to fluid mechanics and many other kinds of engineering disciplines.” He adds that it can be applied to determine various properties, such as magnetic fields in fusion reactors. It is “very universal, not only in different materials, but in different fields”.

Yang says he first started thinking about this approach when he was researching material data where some of the images he was using were blurry. Blurry area. “How can this lost information be recovered?” he wondered. Further reading revealed that this is an example of a broader problem of trying to recover lost information, known as the inverse problem.

Developing this method involved an iterative process of having the model make preliminary predictions, comparing it to the actual data for the material in question, and further fine-tuning the model to match that information. rice field. The resulting model was tested against the case where the materials were well understood to allow calculation of their true internal properties, and the predictions of the new method were in good agreement with the calculated properties.

The training data included not only images of the surface, but also measurements of various types of surface properties such as stress, electric field, and magnetic field. Researchers often used simulated data based on their understanding of the underlying structure of a particular material. And even if the new material has many unknown properties, this method can produce good enough approximations to give engineers general direction on how to proceed with further measurements.

As an example of how this methodology might be applied, Buehler notes that testing an entire airplane is not practical, so a few representative areas are currently examined by expensive methods such as X-rays. It points out that the plane is often inspected by testing. “It’s a different approach, a much cheaper way to collect data and make predictions,” he says. “From there, you can decide where you want to look and test with more expensive equipment.”

First, he expects the method, which is freely available for anyone to use through the GitHub website, to be applied primarily in laboratory settings, such as testing materials used for soft robotics applications. increase.

For such materials, “we can measure things on the surface, but we often don’t know what’s going on inside the material, because materials can be hydrogels, proteins, or actuators.” Because it’s made of biomaterials and there’s no theory, so it’s just that researchers are using our technology to predict what’s going on inside, and maybe better grippers or better composites It’s an area where you can design ,” he adds.

This article was reprinted courtesy of MIT News (web.mit.edu/newsoffice/), a popular site for MIT research, innovation, and education news.