Deriving the stress-strain response of materials is fundamental in the construction industry. Engineers must assess the capability of structures to withstand the design loads and the correspondent deformations. Nevertheless, the stress-strain response of materials can be determined via analytical models of high complexity and when this is not feasible via other computational techniques such as the Finite Element or Finite Difference methods (FEM and DEM).
The difficulties that engineers face in predicting the stress-strain response include the geometry of the model, the chosen constitutive model and the boundary conditions that each problem requires. Hence, it is an issue that requires knowledge, judgement and time. On the contrary, the new method proposed handles the problem using an entirely different approach in which AI is employed to do all that. "[Solving a problem using conventional methods] is very expensive — it can take days, weeks, or even months to run some (computer) simulations. So, we thought: Let’s teach an AI to do this problem for you,” Markus Buehler, co-author of the study and the McAfee Professor of Engineering at MIT, stated.
What the new approach does is utilizing image processing via an AI algorithm to derive the stress and strain of materials. It does not incorporate any of the complex physical models and thus, it is less time-consuming and simpler to use. Of course, the development of the algorithm was not an easy task. Scientists employed a Generative Adversarial Neural Network which was trained using thousands of image sets. Each image set included i) an image showing the internal microstructure of a material subjected to a certain stress field and ii) another color-based image with the resulting stresses and strains of the material in that stress field. Thus, the algorithm was trained to derive the stress-strain field via the geometry of a material's microstructure. Consequently, given an image, the Neural Network developed can predict the forces applied and the resulting deformations of a structural component.
The Machine Learning technique was tested using actual data and was found to be accurate. It could even detect the formation of cracks in areas where the stress-strain field quickly altered.
A major problem in material science and engineering is the scale of interest. The laws and equations that govern the microstructural response of a material do not apply to larger scales such as that of a construction project. Therefore, different approaches and models have been developed depending on the scale of interest. However, the authors of the study suggest that the new tool can readily deal with multiple scales using convolution techniques. “That’s why these neural networks are a great fit for describing material properties,” Prof. Buehler, added.
The researchers focused on composite materials with both ductile and brittle nature. However, more material types will be investigated in the future.
The tool could be used to replace the complex, conventional methods to detect a material's stress-strain response in multiple industries (construction, aviation, automotive) saving time and funds. Moreover, due to its simplicity, it could also be used by professionals that are not experts in this field.
Sources: MIT Yang et al. (2021)
