LLNL-Collaboration / BYU_Alkemi_2017

BYU Big Data Capstone Project (Class 2017)
0 stars 1 forks source link

Applying Machine Learning to Simulated Physics Data

BYU: Jared Hoff, Matthew Oehler, Rebecca Petersen, Andrew Wheeler

LLNL: Brian Gallagher, Cyrus Harrison, Ming Jiang

We invite BYU students to work with LLNL computer scientists to explore how data science techniques can be applied to analyze HPC datasets. LLNL is a leader in physics-based computational simulation and modeling. Our scientists produce massive time varying datasets in support of their work. We are staring to look deeper into these datasets using machine learning.

In this effort you will leverage machine learning and computational statistics methods to analyze feature vectors and computational meshes extracted from simulations in suport of: Simulation Workflow Automation.

LLNL's physics simulation codes provide a complex menu of modeling choices and parameters. Setting up a successful simulation is often a trial-and-error process. The user makes modeling choices, runs the simulation to failure, rolls back the simulation to a valid state, and then adjusts parameters to (hopefully) avoid failure. The cycle can be tedious, which is why the Alkemi project at LLNL is developing ways to help automate aspects of this workflow.

The Alkemi team is applying machine learning on mesh and physics features from simulation data to predict simulation failures ahead of time and automate the roll back and adjustment process. They are also working to extract the characteristics of robust simulation workflows, to help reduce the number of iterations in the process.