A GPU Computational Facility for ML and AI Based Design of Multifunctional Materials
Abstract
This DURIP project will establish a cutting-edge GPU-accelerated computational facility for Machine Learning (ML) research geared towards the design and manufacturing of multifunctional materials. Featuring Dell PowerEdge C4140 computing facility with 16 NVIDIA Tesla V100 GPUs which are linked by NVLink bridge in a shared computing environment and a Dell PowerEdge R740XD storage server with 288 TB capacity, it will vastly accelerate the exploitation of ML in materials modeling, design and synthesis. Following the materials-by-design paradigm, our work will serve DoD needs by providing a fast and comprehensive protocol to develop hierarchical materials with the highest performance under extreme conditions, e.g. high toughness, high flexibility, low density, high resilience, using machine learning methods. The implementation of GPU computing enables faster Artificial Intelligence (AI) techniques that close the loop between theory and experiment in materials design and synthesis, supporting and strengthening several DoD and other federally funded projects. The key distinguishing features of our new facility are: (1) High performance GPU computing: The new GPU will be able to provide an average of 47 times performance boost for ML algorithms compared with Xeon CPUs. In addition to the benefits on ML algorithms, GPU acceleration alongside CPU parallelization can outperform a CPU cluster by 10-20 times for molecular dynamics (MD), coarse-grained (CG), finite element (FEM) and similarsimulations; and (2) High-volume data processing: The proposed GPU cluster with more than 500 GB VRAM memory will provide an extraordinary ability for processing of enormous data sets of 100s of GBs, at the same time. The implementation of GPU acceleration enables faster Artificial Intelligence (AI) techniques that close the loop between theory and experiment in materials design. On the one hand, this GPU cluster will accelerate ML algorithms (e.g. deep convolutional neural networks (Lecun et al., 1998), long short-term memory units (LSTMs) (Hochreiter and Schmidhuber, 1997), and generative adversarial networks (GANs) (Goodfellow et al., 2014), both in training neural networks and generating solutions by using supervised learning and unsupervised learning. APPROVED FOR PUBLIC RELEASE
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Apr 29, 2020
- Source ID
- N000142012189
Entities
People
- Markus J. Buehler
Organizations
- Massachusetts Institute of Technology
- Office of Naval Research
- United States Navy