Translated from: news.mit.edu/2021/ai-mat…
Newton may have met his match.
For centuries, engineers have relied on the laws of physics developed by Newton and others to understand the stresses and strains in the materials they use. But solving these equations can be a computational challenge, especially for complex materials.
MIT researchers have developed a technique that can quickly determine certain properties of a material, such as stress and strain, based on images that show its internal structure. This approach could one day eliminate the need for arduous physics-based calculations and instead rely on computer vision and machine learning to generate estimates in real time.
The researchers say the development could lead to faster prototyping and material inspection. “This is a completely new approach,” Yang said, adding that the algorithm “does the whole thing without any knowledge of the field of physics.”
The study is published today in the journal Science Advances. Zhenze Yang is the lead author of the paper and a doctoral student in the Department of Materials Science and Engineering. Co-authors include Zhihua Yu, a former MIT postdoctoral fellow, and Markus Buehler, the McAfee Professor of Engineering and director of the Atomic and Molecular Mechanics Laboratory.
Engineers spent a lot of time solving equations. They help reveal the internal forces of a material, such as stresses and strains, which can cause the material to deform or break. Such calculations might indicate how a proposed bridge would hold up under heavy traffic or high winds. Unlike Sir Isaac, today’s engineers do not need paper and pen to do this task.” “Many generations of mathematicians and engineers have written down these equations and then figured out how to solve them on computers,” Buehler said. But it is still a thorny issue. It’s very expensive — running some simulations can take days, weeks, even months. So we want to. Let’s teach an ARTIFICIAL intelligence to do this problem for you.”
The researchers turned to a machine learning technique called generative adversarial neural networks. They trained the network with thousands of paired images — one depicting the internal microstructure of a material subjected to mechanical forces, the other depicting the colour-coded stress and strain values of the same material. Armed with these examples, the network used game theory to repeatedly calculate the relationship between the geometry of a material and the stresses it generated.
“So from a picture, the computer can predict all these forces: deformations, stresses, and so on, “Buehler said. This is a real breakthrough — in the traditional way, you code equations and ask a computer to solve partial differential equations. We just go from picture to picture.”
This visualization shows the performance of deep learning methods in predicting physical fields given different input geometry. The image on the left shows the different geometries of the composite, where the soft material is elongating, and the image on the right shows the predicted mechanical field corresponding to the geometries on the left.
This image-based approach is particularly beneficial for complex composites. “Forces on materials may behave differently at the atomic scale than they do at the macroscopic scale.” If you look at an airplane, you probably have glue, metal and polymer in the middle. So, you have all these different dimensions and different scales to determine the solution, “Buller said. If you go the hard way — Newton Road — you have to take a giant detour to get the answer.”
But the researchers’ network is good at handling multiple scales. It processes information through a series of “convolution”, analysing images at progressively larger scales. “That’s why these neural networks are very good for describing material properties,” Buehler says.
The fully trained network performed well in the tests, successfully presenting stress and strain values in a series of close-up images of the microstructure of various soft composites. The network can even pick up “singularities”, such as cracks in materials. In these cases, forces and fields vary rapidly over tiny distances.” “As a materials scientist, you want to know if the model can reproduce these singularities,” Buehler said. And the answer is yes.
This visualization shows the failure of complex materials simulated by a machine learning-based approach without solving the governing equations of mechanics. Red represents soft materials, white represents brittle materials, and green represents cracks.
Suvranu De, a mechanical engineer at Rensselaer Polytechnic Institute who was not involved in the research, said the development could “significantly reduce the number of iterations needed to design a product.” The end-to-end approach proposed in this paper will have a significant impact on a variety of engineering applications — from composites used in the automotive and aircraft industries to natural and engineered biomaterials. It will also have major applications in the field of pure scientific exploration, as force plays a key role in a surprising array of applications, from micro/nano electronics to cell migration and differentiation.”
In addition to saving engineers time and money, the new technology could give non-specialists access to state-of-the-art material calculations. For example, architects or product designers can test the feasibility of their ideas before handing them over to the engineering team. “They can just draw their scenario and figure it out,” Buehler said. This is a big deal.
Once trained, the network can run almost instantaneously on consumer-grade computer processors. This allows mechanics and inspectors to diagnose potential problems with machines simply by taking pictures.
In the new paper, the researchers worked primarily with composite materials, which include a variety of soft and brittle components in random geometric arrangements. In future work, the team plans to use a wider range of material types.” “I really think this approach will have a huge impact, “Buehler said. Empowering engineers with ARTIFICIAL intelligence, that’s really what we’re doing here.
The study was funded in part by the Army Research Office and the Office of Naval Research.