Exploring Gravitational-Wave Detection and Parameter Inference using Deep Learning

The 2022 ITP Postdoctoral Symposium

He Wang (王赫)

co-advisor: Zong-Kuan Guo (郭宗宽)

hewang@mail.bnu.edu.cn / hewang@itp.ac.cn

hewang@itp.ac.cn / hewang@mail.bnu.edu.cn

FYI: I will speak in Chinese for the sake of clarity.

• Motivation
• GW detection
• GW parameter estimation
• In progress...

Content

Motivation

• It's necessary to speed up...
• It has the potential to discover more...

• The number of papers on ML applications to GW data has grown rapidly in recent years (See Cuoco, et al Mach. Learn.: Sci. Technol (2020) for a review or site Survey4GWML)
• The most recent and exciting work has been on rapid parameter estimation

Gravitation wave data analysis with machine learning

From: LIGO-G2102497

Binary detection rates

• O3 ~ 1/5 days
• O4 ~ 1/2 days
• O5 ~ 3 / day

Simulated Event Stream for a one year duration O4 run

• The current Matched filtering technique is computationally expansive.
• More GW events are coming...

• We proposed a brand-new architecture, called matched-filtering convolutional neural network (MFCNN), as a GW trigger generator
• Advantages combined:
  • Matched filtering (weak signal extraction) + AI (pattern recognition)
• Earth-based GW detection
  • Recovering the three GW events in O1.
  • Recovering all GW events in O2, even including GW170817 event.
  • With 33 events (H1+L1) in O3a (39 events in total), it has 87.9% sensitivity detection rate, only 4 of them are misclassified.
  • Efficiency: ~ ms
• Although it has an extraordinary sensitivity to GW detection, the FAR is embarrassing. (FAR: 1/ day)

Ref: He Wang, Shichao Wu, Zhoujian Cao, Xiaolin Liu, and Jian-Yang Zhu. Physical Review D 101, no. 10 (May 2020): 104003.

GW detection

• Space-based GW detection (MBHBs)
  • We present a model with robust sensitivity to numerous GW sources and modulation of massive black hole binaries (MBHBs) waveform family.
  • Our model is capable of processing 1-year of data, simulated from the LISA data challenge (LDC), in several seconds only, while identifying all coalescences of MBHBs with no false alarms.
  • Our analysis represents a starting point from which a neural network trained on Gaussian and stationary background noise can be applied to realistic non-Gaussian and non-stationary data.
  • The generalization ability of the supervised learning approach can be extended to various TDI configurations and can also be greatly useful for future space-based GW detectors.
• Again it has an extraordinary sensitivity to GW detection, but we still don't know how to design a FAR for each candidates.

Ref: Wen-Hong Ruan, He Wang, Chang Liu, and Zong-Kuan Guo. ArXiv Preprint ArXiv:2111.14546, November 2021.

GW detection

Ref: CunLiang Ma, Wei Wang, He Wang, and Zhoujian Cao. Physical Review D 105, no. 8 (April 25, 2022): 083013.

• Q: How to produce the detection statistics based on ?
• A: Try ensemble learning!
   
• An ensemble model is capable of processing 1-month of LIGO data from O2 identifying all coalescences of CBC with no false alarms.
• FAR could be reduced dramatically with the ensemble learning method. 
• FAR: 1 / 0.3 month

GW detection

• Curse of Dimensionality
  • May take several days to several weeks to run parameter estimation on a single event.

  • Traditional statistical methods scale poorly in time/accuracy for
    datasets described by many parameters.

  • Machine learning algorithms perform very well in exploiting
    correlations across a large number of dimensions

  • We tackle this issue by thinking out of the box and directly using the information from the "future", as opposed to the others' work by exploring network structure.

• “Prior Sampling”

  • Therefore the key question is how to effectively sample the feature space.

  • This is essentially equivalent to incorporating the physical domain knowledge into the high-dimensional training data.

  • In our case, we use the interim distribution (alpha=1) that is derived from the Monte Carlo method as a representation of the prior physical knowledge.

A diagram of prior sampling between feature space and physical parameter space

Parameter estimation

Ref: He Wang, Zhoujian Cao, Yue Zhou, Zong-Kuan Guo, and Zhixiang Ren. Big Data Mining and Analytics 5, no. 1 (2022): 53–63.

• It implies that roughly 10% of physical prior knowledge incorporated is enough for accurate Bayesian inference of the high-dimensional gravitational-wave data.
• Although we recognize a similar idea that applies the normalizing flow technique to characterize the distribution exits in the previous works [Green & Gair (2020), 2106.12594], our method still differs from the recent study in GW study:
  1. Improved data preparation
     (sampling all 15 dims instead of extrinsic only)
  2. Prior sampling to construct the training dataset with domain knowledge using the SMOTETomek technique.
     (Relatively a better performance than [Green & Gair (2020)])
  3. Effectively and efficiently perform Bayesian inference on ultra-high dimension gravitational-wave data.
     (8s for 50,000 posterior samples)

Parameter estimation

Ref: He Wang, Zhoujian Cao, Yue Zhou, Zong-Kuan Guo, and Zhixiang Ren. Big Data Mining and Analytics 5, no. 1 (2022): 53–63.