Modeling, detection and characterization of eccentric compact binary coalescence

Transcription

1 Modeling, detection and characterization of eccentric compact binary coalescence E.A. Huerta et al., Phys. Rev. D 95, (2017) Daniel George, E.A. Huerta, arxiv: E.A. Huerta et al., in preparation Eliu Huerta National Center for Supercomputing Applications (NCSA) University of Illinois at Urbana-Champaign NCSA: Bhanu Agarwal, Daniel George, Roland Haas, Miguel Holgado, Diyu Luo, Wei Ren, Erik Wessel Cambridge: Chris Moore and Alvin Chua CITA: Prayush Kumar and Harald Pfeiffer AEI: Ian Hinder Aspen Center for Astrophysics, The Dawning Era of Gravitational-Wave Astrophysics, February 2017

2 Outline Motivation State-of-the-art in inspiral-merger-ringdown (IMR) eccentric compact binary coalescence (ecbc) source modeling Second generation of IMR ecbc waveform models: combining analytical and numerical relativity with machine learning Detection and characterization of ecbc: introducing a new paradigm in gravitational wave astrophysics

3 Motivation Advanced LIGO (aligo) has established itself as an astronomical observatory: the dark sector of the Universe is at our fingertips Enable discovery, expand the science reach of gravitational wave astrophysics by extracting as much information as possible from aligo data Nature is oblivious to our assumptions. Innovate to confirm/rule out the existence of ecbc populations (Carl Rodriguez, Bence Kocsis, Zsolt Frei) Gravitational wave astrophysics informs astrophysical theory current detections have reshaped our expectations and knowledge of compact binary populations

4 State-of-the-art E. A. Huerta et al., Phys. Rev. D 95, (2017) introduced the first eccentric IMR waveform model that captures the effect of orbital eccentricity throughout the merger of eccentric, non-spinning compact binary sources. Key features: Reproduces state-of-the-art quasi-circular waveform models in the zero eccentricity limit Captures the effect of orbital eccentricity throughout the merger phase Ready-to-use for large scale data analysis studies

5 E.A. Huerta et al., Phys. Rev. D 95, (2017) Hybrid Inspiral Framework Conservative Dynamics Derived from a 3PN Hamiltonian Incorporate quasi-circular selfforce corrections to the binding energy of compact binaries (up to 6PN order) Recast expressions using a gauge-invariant expansion parameter Radiative Dynamics Energy and angular momentum fluxes Derive higher-order eccentric PN corrections both for instantaneous and hereditary terms Incorporate higher-order quasi-circular corrections to the fluxes from black hole perturbation theory (up to 6PN order)

6 E.A. Huerta et al., Phys. Rev. D 95, (2017) Merger waveform model Assume that moderately eccentric systems circularize prior to the merger event e 0 at f GW = 15Hz

7 E.A. Huerta et al., Phys. Rev. D 95, (2017) Merger waveform model Catalog of NR simulations to calibrate a phenomenological merger waveform Generic Implicit Rotating Source (girs) Model

8 E.A. Huerta et al., Phys. Rev. D 95, (2017) Hybrid Inspiral Scheme Analytical Merger Waveform Overlap between e! 0 IMR ax templates and non-spinning SEOBNRv2 templates. Overlap between IMR ax templates and non-spinning SEOBNRv2 templates. e 0 at f GW = 14Hz. Filtering from f GW = 15Hz aligo s design sensitivity.

9 DRAFT DRAFT DRAFT

10 DRAFT DRAFT DRAFT

11 E.A. Huerta et al., Phys. Rev. D 95, (2017) The effect of orbital eccentricity in amplitude and phase is captured throughout the merger \

12 Detection of ecbc E.A. Huerta et al., Phys. Rev. D 95, (2017) Fitting Factor (FF) calculations 1.5M SEOBNRv2 spin-aligned templates e 0 at f GW = 14Hz. Filtering from f GW = 15Hz using aligo s design sensitivity.

13 Detection of ecbc E.A. Huerta et al., Phys. Rev. D 95, (2017) Fitting Factor (FF) calculations 1.5M SEOBNRv2 spin-aligned templates 1.5M templates to e 0 at f GW = 14Hz. get convergent Filtering from f GW = 15Hz using aligo s design sensitivity. results!!! Unfeasible in low latency

14 State-of-the-art Key areas of improvement for ecbc waveform modeling: Merger evolution Amplitude evolution Improve accuracy of IMR ecbc model in the quasi-circular limit: waveform inaccuracies vs effect of eccentricity for nearly quasi-circular CBC

15 E. A. Huerta et al., in preparation Second generation of IMR ecbc waveforms Use hybrid inspiral framework from Phys. Rev. D 95, (2017) Improve modeling of amplitude evolution by including higher-order PN corrections Develop a new merger waveform using machine learning

16 E. A. Huerta et al., in preparation Gaussian Process Regression Gaussian Process Regression (GPR) has been used by Cambridge University s gravitational wave team in the context of source modeling and parameter estimation Moore et al PRL 113, (2014) Moore et al Phys. Rev. D 93, (2016) See Dan Holz s group work on GPR. We have completed the construction of a GPR merger waveform and linked it to our hybrid inspiral waveform

17 State-of-the-art E.A. Huerta et al., Phys. Rev. D 95, (2017) Hybrid Inspiral Scheme Analytical Merger Waveform While the implicit rotating source (IRS) merger waveform model provides a good description of the merger dynamics in the vicinity of the light-ring, its accuracy deteriorates rapidly several cycles before merger (REPLACE IRS-based merger waveform!) The corrections that have to be implemented to ensure that the minimum overlap between the IMR ax model and SEOBNRv2 are greater than 0.99 over the whole BBH space are within reach with additional work (this talk!) Overlap between e! 0 IMR ax templates and non-spinning SEOBNRv2 templates. Filtering starts at f GW = 15Hz using design sensitivity for aligo. E.A. Huerta et al., Phys. Rev. D 95, (2017)

18 Comparison to state-of-the-art quasi-circular waveforms E. A. Huerta et al., in preparation No free parameters used for the construction of the model Combination of PN, self-force, black hole perturbation theory and numerical relativity GPE merger waveform constructed with numerical simulations Inspiral evolution was combined with GPR merger waveform Hybrid inspiral framework encodes the correct physics very late in the inspiral evolution DRAFT

19 E. A. Huerta et al., in preparation NCSA-CAM ecbc waveform The inspiral dynamics of non-spinning, quasi-circular compact binaries can be accurately described without resorting to phenomenology no free parameters or re-summations needed Non-spinning, quasi-circular IMR dynamics is accurately captured by combining our hybrid inspiral framework with a small training set of numerical relativity waveforms and machine learning Our new NCSA-CAM also encapsulates the effect of orbital eccentricity throughout the merger

20 Validation of NCSA-CAM ecbc waveform with eccentric NR simulations E. A. Huerta et al., in preparation Uncalibrated NCSA-CAM model compared to eccentric NR simulations DRAFT DRAFT

21 Now what? Basic analysis requires 1.5M+ templates Extension to spinning, eccentric compact binaries? What is the status of searches for quasi-circular, spin-aligned binaries? 300k+ templates needed Is this model sustainable? Open Science Grid up and running in Blue Waters this can only go so far. ATLAS, NANOGrav taking advantage of this platform NR and EM followups require rapid and accurate PE results to enable real-time multimessenger astrophysics

22 Now what? Matched-filtering searches are computationally expensive Limited by the number of templates needed to carry out the search Extension to explore a deeper parameter space is computationally prohibitive Are we missing astrophysically motivated sources lurking in aligo data Advanced Virgo, KAGRA and LIGO-India will eventually come on-line Do we go and seize all HPC and HTC resources to detect and characterize new GW sources?

23 Deep Neural Networks RCATJLTHYHATOJ

24 Deep Neural Networks RCATJLTHYHATOJ

25 Deep Neural Networks Daniel George, E. A. Huerta, arxiv: First scientific application for processing highly noisy time data series Advantages Directly processes raw data Automatically develops optimal strategy Constant evaluation time no matter the size of the training dataset Highly non-linear Optimized hardware for deep learning (GPU) Features Does not perform template matching (linear filter) Learns patters connecting signals Dimensional reduction Interpolation Compact, portable code

26 Deep Neural Networks Daniel George, E. A. Huerta, arxiv: First scientific application for processing highly noisy time data series

27 Deep Neural Networks Daniel George, E. A. Huerta, arxiv: First scientific application for processing highly noisy time data series Design of DNN Used Wolfram Language MXNet Innovative Systems Lab at NCSA: Tesla, GTX1080 and P100 GPUs Basic design: 4000 templates Fully trained DNN is 4MB in size

28 Deep Neural Networks Daniel George, E. A. Huerta, arxiv: First scientific application for processing highly noisy time data series Parameter Estimation

29 Daniel George, E. A. Huerta, arxiv: Application to ecbc DRAFT

30 Conclusions New approach to combine an accurate hybrid inspiral framework with a machine learning-based merger waveform We have introduced the second generation of IMR, non-spinning, ecbc waveforms: no free parameters or re-summations needed. Validated with state-of-the-art quasi-circular waveforms and eccentric NR simulations Detection and characterization of moderately eccentric compact sources is now within reach with deep neural networks stay tuned!

31 ecbc Team

32 Acknowledgements This research is part of the Blue Waters sustained-petascale computing project, which is supported by the National Science Foundation (awards OCI and ACI ) and the state of Illinois. Blue Waters is a joint effort of the University of Illinois at Urbana-Champaign and its National Center for Supercomputing Applications. The eccentric numerical relativity simulations used in this article were generated with the open source, community software the Einstein Toolkit. Vlad Kindratenko for granting us access to several NVIDIA Tesla, GeForce, and P100 GPUs in the Innovative Systems Lab at NCSA. Wolfram Research for developing the software stack used for DNN analysis.