Datasets for Machine Learning
Phase-field modeling can be applied to predict the responses of polycrystalline materials to external stimuli (including for instance magnetic fields, electric fields, and stress), thereby enabling the predicting of their magnetic, electric, and mechanical properties, respectively. Phase-field model can also be used to predict responses/properties based on coupling effects, such as piezoelectric (the development of charges in response to stress) and magnetostrictive coupling (the development of strain in response to magnetic fields). However, direct numerical simulations of 3D polycrystalline systems is usually quite time-consuming, especially for large-scale (microns or larger) systems.
To address this issue, we, on one hand, are actively developing advanced numerical and parallelization algorithms, to enable large-scale phase-field modeling via massive parallelization; on the other hand, exploiting the state-of-the-art deep learning tools.
Here, we set up the datasets composed of different polygrain structure and its corresponding properties. We use these data to train Graph Neural Network (GNN), a recent development in the field of deep learning, for a fast (x1000 or more faster than phase-field model and other direct numerical simulations methods), accurate, and interpretable prediction of the materials properties.
This project is in collaboration with Prof. Yingyu Liang and Mehmet Furkan Demirel from Department of Computer Science at the University of Wisconsin-Madison.
To address this issue, we, on one hand, are actively developing advanced numerical and parallelization algorithms, to enable large-scale phase-field modeling via massive parallelization; on the other hand, exploiting the state-of-the-art deep learning tools.
Here, we set up the datasets composed of different polygrain structure and its corresponding properties. We use these data to train Graph Neural Network (GNN), a recent development in the field of deep learning, for a fast (x1000 or more faster than phase-field model and other direct numerical simulations methods), accurate, and interpretable prediction of the materials properties.
This project is in collaboration with Prof. Yingyu Liang and Mehmet Furkan Demirel from Department of Computer Science at the University of Wisconsin-Madison.
Different 3D microstructures is generated from grain statistics (size distribution, shape ratio, texture, etc.) using Dream.3D. Magnetostrictive response of each generated microstructure is predicted using phase-field modeling.
Datasets can be downloaded here. The zipped folder includes 492 subfolder corresponding to 2287 different data points. We are still working to enlarge the datasets. |