SHARE
Research Highlight

HydraGNN - Distributed PyTorch Implementation of Multi-Headed Graph Convolutional Neural Networks

Topic:

Brief: Researchers from the Computing and Computational Sciences Directorate (CCSD) at Oak Ridge National Laboratory (ORNL) have developed a distributed implementation of graph convolutional neural networks [1]. The code has been shown to successfully take advantage of high-performance computing (HPC) resources on ORNL-CADES-DGX, ORNL-CADES-CONDO, and ORNL-OLCF-SUMMIT to produce fast and accurate predictions of macroscopic material properties using atomic information.

Accomplishment: Researchers from CCSD at ORNL have developed a distributed PyTorch implementation of multi-headed graph convolutional neural networks (GCNNs) to produce fast and accurate predictions of graph properties [2]. The Artificial Intelligence for Scientific Discovery (AISD) Thrust of the Artificial Intelligence Initiative at ORNL has funded this effort to produce fast and accurate predictions of material properties using atomic information. Within this context, GCNNs abstract the lattice structure of a solid material as a graph, whereby atoms are modeled as nodes and metallic bonds as edges. This representation naturally incorporates information about the structure of the material, thereby eliminating the need for computationally expensive data pre-processing which is required in standard neural network (NN) approaches. We trained GCNNs on ab-initio density functional theory (DFT) for copper-gold (CuAu) [2] and iron-platinum (FePt) [3] data generated by running the LSMS-3 code, which implements a locally self-consistent multiple scattering method, on Oak Ridge Leadership Computing Facility (OLCF) supercomputers Titan and Summit. GCNN outperforms the ab-initio DFT simulation by orders of magnitude in terms of computational time to produce the estimate of the total energy, charge transfer, and magnetic moment for a given atomic configuration of the lattice structure. We compared the predictive performance of GCNN models against a standard NN such as dense feedforward multi-layer perceptron (MLP) by using the root-mean-squared errors to quantify their predictive quality. We found that the attainable accuracy of GCNNs is at least an order of magnitude better than that of the MLP.  

Albeit the code has been originally developed in response to the scientific needs addressed by the AISD Thrust of the AI Initiative, its framework is general enough to support other scientific applications at ORNL such as transportation, power grid, and cybersecurity. 

HydraGNN - Distributed PyTorch Implementation of Multi-Headed Graph Convolutional Neural Networks ORNL Oak Ridge National Laboratory CCSD AI Initiative
HydraGNN predictions on test portion of FePt datatest set. Scatter plots of mixing enthalpy
(left), atomic charge transfer (center), and atomic magnetic moment (right) for multi-task predictions obtained with HydraGNN (y-axis) against DFT calculations (x-axis). The color map indicates the relative frequency of data for each predicted property.

Acknowledgement: This research was funded by the AI Initiative, as part of the Laboratory Directed Research and Development Program of Oak Ridge National Laboratory, managed by UT-Battelle, LLC, for the U.S. Department of Energy (DOE). This research used resources of the Oak Ridge Leadership Computing Facility (OLCF) at the Oak Ridge National Laboratory, which is supported by the Office of Science of the U.S. Department of Energy under Contract No. DE-AC05-00OR22725.

Publications and presentations resulting from this work:

  1. Lupo Pasini, M., Zhang, P., Reeve, S.T., Choi, J.Y, Multi-task graph neural networks for simultaneous prediction of global and atomic properties in ferromagnetic systems, ArXiv Pre-print https://arxiv.org/abs/2202.01954, Submitted to peer-reviewed journal.
  2. Zhang, P., Reeve, S.T., Choi, J.Y., Lupo Pasini, M., Exascale-capable graph convolutional neural network surrogates for atomic property prediction, Presentation to the TMS Annual Meeting 2022 - Algorithm Development in Materials Science and Engineering: ML Algorithms and Their Applications.
  3. Lupo Pasini, M., Zhang, P., Reeve, S.T., Choi, J.Y., Graph convolutional neural networks for fast, accurate prediction of material properties for solid solution high entropy alloys using open-source datasets, Presentation to the TMS Annual Meeting 2022 - ICME Case Studies: Successes and Challenges for Generation, Distribution, and Use of Public/Pre-Existing Materials Datasets mini-symposium.
  4. Lupo Pasini, M., Deep learning for prediction of material properties of solid solution alloys from multiscale information, ORNL Computational Mechanics Seminar Series, https://computmech.ornl.gov/seminars.html, Thursday | February 24 | 2 p.m.–3 p.m.
  5. Lupo Pasini, M., Fast and accurate predictions of material properties from atomic information using graph convolutional neural networks, 34th Annual CSP Workshop, Recent Developments in Computer Simulation Studies in Condensed Matter Physics February 22 - 25, 2022, Invited by David Landau.

Technology Transfer resulting from this work:

  1. Lupo Pasini, Massimiliano, Reeve, Samuel T., Zhang, Pei, and Choi, Jong Youl. HydraGNN. Computer Software. https://github.com/ORNL/HydraGNN. USDOE. 19 Oct. 2021. Web. doi:10.11578/dc.20211019.2.

Contact: Massimiliano Lupo Pasini (lupopasinim@ornl.gov) 
Team:   Massimiliano (Max) Lupo Pasini, Samuel Temple Reeve, Pei Zhang, Jong Youl Choi

References: 

  1. Lupo Pasini, Massimiliano, Reeve, Samuel T., Zhang, Pei, and Choi, Jong
  2. Youl. HydraGNN. Computer Software. https://github.com/ORNL/HydraGNN. USDOE. 19 Oct. 2021. Web. doi:10.11578/dc.20211019.2.
  3. Lupo Pasini, Massimiliano, Zhang, Pei, Reeve, Samuel T., and Choi, Jong Youl. Multi-task graph neural networks for simultaneous prediction of global and atomic properties in ferromagnetic systems, ArXiv preprint, arXiv:2202.01954
  4. Lupo Pasini, Massimiliano and Eisenbach, Markus. CuAu binary alloy with 32 atoms - LSMS-3 data. United States: N. p., 2021. Web. doi:10.13139/OLCF/1765349.
  5. Lupo Pasini, Massimiliano and Eisenbach, Markus. FePt binary alloy with 32 atoms - LSMS-3 data. United States: N. p., 2021. Web. doi:10.13139/OLCF/1762742.