Fracture forecasting with deep neural networks
Fracture forecasting with deep neural networks lead image
Understanding how materials crack, and how those cracks evolve, is important for understanding how materials behave under pressure and for developing new materials to withstand mechanical stress. However, fracturing is a complex process, and simulating it on the atomic scale is difficult and expensive.
To circumvent the limitations imposed by conventional simulations, Hsu and Buehler developed DyFraNet, which uses a deep neural network to model how cracks develop in materials.
Many parameters influence fracturing, including the structure of a material and its mechanical properties, the applied mechanical load, and boundary conditions.
“A fracture event is the result of the integrated response of a large number of atoms,” said author Markus J. Buehler. “Their collective behavior across space and time determines the ultimate failure process. Machine learning algorithms can be used to efficiently capture these mechanisms, and they allow us to make predictions about these collective, integrated atomic behaviors.”
DyFraNet uses 2D images that encode the features that contribute to the fracture event and calculates how the crack may evolve.
“The key to this physics-inspired neural framework is following the relationship between physical parameters and resulting phenomena, including an explicit representation of space and time to capture the complex statistical behaviors of bond breaking that govern fracture. The integration of space and time allows us to build a rigorous relationship between physical parameters and dynamical results in our machine learning model,” said Buehler.
The authors believe their machine learning framework could be applied to study other dynamical phenomena. As machine learning helps reveal how materials behave under pressure, it can enable more crack-resistant materials.
Source: “DyFraNet: Forecasting and backcasting dynamic fracture mechanics in space and time using a 2D-to-3D deep neural network,” by Yu-Chuan Hsu and Markus J. Buehler, APL Machine Learning (2023). The article can be accessed at https://doi.org/10.1063/5.0135015 .
