Hybrid Physics-Data Models for Simulating Material Behavior Across the Scales
Speaker: Iuri Rocha
Abstract: We increasingly rely on high-performance materials and structures for a wide range of applications, from wind turbines to medical implants. Yet, due to their highly-optimized nature, these materials have also become increasingly difficult to design and study using experimental techniques alone. In my talk I will go through a few of my group’s recent developments on computational modeling of material across the scales. We will quickly come to the conclusion that the models are so computationally expensive that designing new materials would take decades of simulations. I will present how we use machine learning to accelerate these models, including hybrid models with intact physics-based components directly embedded in neural network architectures. I will then wrap up with a GNN-based approach we recently developed for microscale simulation of material behavior, and show how it fits within our broader vision of hybrid models for computational mechanics.
[ Slides ]
Predicting Plastic Strain Localization in Porous Solids Using Graph Neural Networks
Speaker: Joep Storm
Abstract: The mechanical failure of plastic porous solids is dominated by so-called shear bands between voids. This observation inspired prior research to represent the failure as a Delaunay graph, allowing for an efficient upper-bound analysis that avoids the more complex finite element simulations. In this talk, I will discuss a Graph Neural Network-based approach built around the same principle that aims to improve the accuracy of the analysis. A fully data-driven approach and a hybrid model that utilizes the mechanistic prior are compared on their data requirements and extrapolation ability.
[ Slides ]