Intelligent Middleware, Making CFD Accessible Through Leadership-Scale Deep Learning Research

Project Description

Intelligent ‘Middleware’ designed to achieve the next leap in human-computer productivity. Focused on engineering software such as computational fluid dynamics and manufacturing applications, the end goal of the microsurgeonbot technology development endeavor is the ability for someone without specialized knowledge to be able to operate the system and achieve useful outcomes. Making the capabilities of high performance computing accessible to a larger percentage of workers in design, analysis, and advanced manufacturing, will empower users in a manner that is expected to be transformational.In this Director’s Discretionary project we will be focused on generating test training databases using the Numeca CFD codes (previously demonstrated in our past publications to scale on Titan), deployment of TensorFlow (recently demonstrated by others such as the OLCF’s Mike Matheson on OLCF systems including Titan), and testing the training of deep and wide networks developed in TensorFlow, and testing and reporting on the deployment and parallel scaling strategies (e.g., MPI with model or data parallelism) and subsequent training performance. We have a team of four working on this endeavor with extensive backgrounds in CFD, HPC, and machine learning.