AI Model Uses Photos to Predict Materials’ Stress and Strain

Photo credit: Mike MacKenzie, CC BY 2.0

Engineers use time-tested, evidence-based physical laws to determine how materials will behave in a particular situation. With knowledge of a material’s structural makeup, engineers can calculate the integrity of a design—built or theoretical—to ensure that structures fulfill the functions required of them.

The design innovations of previous decades have led to a rise in the use of composite materials throughout the engineering field. Composite products like concrete, plywood, and fiberglass confront consumers everyday as mainstays of modern design. Composite materials can offer expanded functionalities like increased strength or lightness as compared to their constituent materials. But as the complexity of material resources increases, so too does the complexity of equations required to calculate stresses and strains. Even with the advent of artificial intelligence (AI), up until now, engineers have been forced to code stress and strain equations into networks before AI can generate simulations and solutions.

New research published by Zhenze Yang, Chi-Hua Yu, and Markus Buehler of the Massachusetts Institute of Technology (MIT) reveals a process that can calculate the properties of a material through the use of machine learning and computer vision, rather than the input of differential equations (as is presently required).

Researchers selected a machine learning framework known as a Generative Adversarial Neural Network as the foundation of their AI model. In order to train the network, the team paired images depicting the internal microstructures of various materials under stress with color-coded images of the materials’ stress and strain values. After exposure to thousands of paired images, the network learned to calculate stresses based on the geometry of a material’s structural makeup.

Through extensive testing and AI exposure to additional scenarios, researchers also determined that their network could capture “singularities,” such as developing cracks in concrete, and accurately simulate the force and field changes resulting from such events. Overall, the research reveals a system capable of generating stress and strain calculations with less time, resources, and manpower than any other method known to the field of engineering.

Yang, Yu, and Buehler predict that their approach will lead to faster progressions through the engineering design process. Professionals like architects and materials inspectors will benefit from a tool capable of calculating material integrity with nothing more than a snapshot. In addition, nonexperts will be able to gather materials calculations for small scale and pet projects alike, because a fully trained version of the researchers’ network runs on computers with consumer-grade CPUs (central processing units).
​
To view Yang, Yu, and Buehler’s recent publication in the journal Science Advances, click here.