Massachusetts Institute of Technology (MIT) researchers have developed a deep learning model to estimate the stresses and strains on materials from their images.
“Our end-to-end deep learning model predicts physical fields like stress or strain directly from the material microstructure geometry, and reaches an astonishing accuracy not only for predicted field data but also for derivative material property predictions,” the researchers said.
For this project, the researchers have worked with composite materials, including soft and hard components in various random geometrical arrangements.
For decades, engineers have relied on physical laws to understand the stresses and strains on materials. At an industrial scale, running simulations using computer-aided engineering (CAE) software to gauge the strength of materials is time-consuming and costly.
MIT researchers have built a computer vision and machine learning technique to calculate properties of a material from its image in quick time. Zhenze Yang, a PhD student in the department of material science and engineering at MIT, led the project. He said the new approach could enable faster design prototyping and material inspections.
How does it work
The paper, Deep Learning Model to Predict Complex Stress and Strain Fields in Hierarchical Composites,’ by Yang, and Chi-Hua Yu and Markus J Buehler, explained how to use Generative Adversarial Neural Network or GANs and convolutional neural networks or CNNS to solve complex material/engineering problems.
The MIT researchers trained the network with thousands of paired images — one showcasing a material’s internal microstructure subject to mechanical forces, while the other depicting the same material’s colour-coded stress and strain values. Using game theory, the network iteratively figured out the relationship between the geometry of a material and its resulting stresses.
The computer can predict deformations, stresses, strains, etc. “That’s the breakthrough. Otherwise, you would need to code the equations and ask the computer to solve partial differential equations,” said Buehler.
For example, the image below showcases the deep learning approach in predicting physical fields, given different input geometrics. The left-side graphic shows a varying geometry of the composite in which the soft material is expanding. In contrast, the right-side figure highlights the predicted mechanical field corresponding to the geometry in the left figure.
Researchers at Facebook AI and Google have also developed machine learning techniques to solve advanced mathematical equations such as integration, first-order and second-order differential equations and partial differential equations for various applications.
Citing aeroplanes, Buehler said there are multiple materials like glue, metal, polymer etc. It becomes highly complex to solve them using existing methods as they have various parameters, scales and factors in determining the solution. “If you go the hard way — Newton way — you have to walk a huge detour to get to the answer,” said Buehler.
MIT researchers claimed that their network is adept at dealing with multiple parameters. It processes information through a series of ‘convolutions,’ which analyses the image at large scales. That’s why these neural networks are a perfect fit for describing material properties, said Buehler.
The researchers claimed that its fully trained model rendered successful stress and strain results using a series of close-up images of the microstructure of various soft composite material. Also, the network was able to capture the micro details and singularities like cracks and other deformities.
The graphic below shows the simulated failures in a material by a machine-learning based approach without solving governing equations of mechanics. Red represents a soft material, white depicts a fragile material, and the green represents a crack.
MIT researchers said the technique saves time and money and also give nonexperts access to material calculations. For instance, architectures or product designers can test the feasibility of their ideas before passing the project along to an engineering team. “That’s a big deal,” said Buehler.
Further, mechanics and inspectors across manufacturing, aerospace and other industries can diagnose potential problems using this technique by simply taking a picture of the material they are inspecting. Once the model is trained, the network can run instantaneously on consumer-grade computer processors.
Subscribe to our Newsletter
Get the latest updates and relevant offers by sharing your email.
Join Our Telegram Group. Be part of an engaging online community. Join Here.
Credit: Google News