| About this Abstract |
| Meeting |
2022 TMS Annual Meeting & Exhibition
|
| Symposium
|
Algorithm Development in Materials Science and Engineering
|
| Presentation Title |
Exascale-capable Graph Convolutional Neural Network Surrogates for Atomic Property Prediction |
| Author(s) |
Pei Zhang, Sam Reeve, Max Lupo Pasini |
| On-Site Speaker (Planned) |
Pei Zhang |
| Abstract Scope |
While density functional theory (DFT) is a critical tool in predictive understanding of materials and widely applicable, it is relatively slow and computationally expensive. As an attractive alternative, graph convolutional neural networks (GCNN) have demonstrated significant predictive power as surrogate models for atomistic properties. Despite their success, large network sizes and computational complexity make the training of the GCNNs challenging and potentially unstable. In this work, we develop an accelerated and stable GCNN application built for effective use of heterogeneous hardware and with physical constraints-based regularization. Specifically, we develop a GCNN with a focus on distributed, GPU-capable training and prediction, for ultimate use on the Summit and upcoming exascale Frontier supercomputers, which is based on PyTorch with multiple graph interaction models and multi-headed network architectures. We will present GCNN predictions for open source DFT data sets, contrasting accuracy, stability, and computational cost. |
| Proceedings Inclusion? |
Planned: |
| Keywords |
Computational Materials Science & Engineering, Machine Learning, |