DEEP LEARNING TO ENABLE REAL-TIME GRAVITATIONAL WAVE AND - - PowerPoint PPT Presentation

GPU Technology Conference Silicon Valley, May 8-11, 2017

DEEP LEARNING TO ENABLE REAL-TIME GRAVITATIONAL WAVE AND MULTIMESSENGER ASTROPHYSICS

Daniel George Eliu Huerta Gravity Group National center for supercomputing applications university of Illinois at Urbana-champaign

Enlightened by light and astroparticles

(C) ESA (C) NASA (C) LHC (C) IMGRES (C) SWIFT (C) ICE CUBE (C) NASA

What about events that do not emit light?

From darkness to sound

Numerically simulate extreme astrophysical environments using supercomputers

Some of the most fascinating astrophysical phenomena are governed by strong gravitational interactions

(C) BBC

(C) NCSA

From darkness to sound

LIGO Livingston LIGO Handford

The most sensitive detectors ever constructed by humankind

https://www.kaistaats.com/film/ligo-detection/

From darkness to sound

Detecting gravitational waves with LIGO requires measuring the distance between the Earth and Proxima Centauri with a precession better than a few microns!

4.24 light years

XXI century physics The discovery of the century

(C) NCSA

XXI century physics The discovery of the century

There is much more than collisions of black holes Collisions of neutron stars and black hole-neutron stars are expected to produce electromagnetic and astroparticle counterparts: the holy grail of multimessenger astrophysics

(C) NCSA Gravity Group

What have we accomplished?

Gravitational waves have been detected Confirmed: binary systems of black holes form and coalesce within the age of the Universe Worldwide effort brought together a rich ecosystem of scientists: experimental and theoretical physicists, computer scientists, HPC, HTC, OSG, data analysts, outreach

Scientific Discovery

Gµν = 8π Tµν

Detected: binary black holes Future: supernovae collapse, gamma-ray bursts, oscillating neutron stars…

Models and simulations Theory Observations

What is happening now?

More detectors, more data, more opportunities, more resources (?) Kilometer scale multidetector network Longer gravitational wave detection campaigns More sensitive detectors

(C) LIGO

Computational challenges

Network throughput of gravitational wave data transfer is 1MB/sec. Raw data are a factor 10-30 larger Gravitational wave searches currently target a narrow class of astrophysical events (3D). Increasing the depth of existing searches is computationally prohibitive (8D) Some types of gravitational wave signals go unnoticed with existing detection algorithms

Computational challenges

2018 2013 2014 2015 2016 2017 2019 2020

Stampede/UT Austin Blue Waters/UIUC Comet/UCSD Wrangler/UT Austin Bridges/CMU/PSC Jetstream/Indiana U. Yellowstone/NCAR-Wyoming

Key: Blue: Large-scale computation Red: Long-tail and high- throughput Green: Data Intensive Orange: Cloud

Current and future portfolio of NSF-supported National Computing Resources

Complements Larger Aggregate Investments from Universities and other Agencies

Stampede2/UT Austin Cheyenne/NCAR 2 to 3x Time-To-Solution Improvement

2021 2022 2023 2024 2025

> 20x Improvement

2026 ...

Leadership HPC Planning

National HPC Resource National HPC Resource Potential Renewal

Waveform templates: minutes to years Waveform templates: seconds to minutes

It is time to connect our successful present with a bright future

Current challenges can only be overcome through innovation Leverage existing HPC and HTC infrastructure with recent breakthroughs in artificial intelligence