重力波を機械学習(DINGO)で解析したい!(前半)/重力波の背景ノイズを解析してみた(後半)

スライド概要

前半では，機械学習によるパラメータ推定ソフト「DINGO-IS」をNVIDIA RTX A5000を搭載したコンピュータに構築し，結果の再現を試みた．代表例としてGW190412(合体した天体の質量が大きく異なる)とGW190828(合体した天体の質量が近い)の結果を紹介している．DINGOの推論がうまく行っているかどうかの指標となるIS効率は，前者で0.01%，後者で約3%となり，特に後者については，DINGOの公式論文の結果(10%台)とは大きく異なる．
後半では文献[5]を参考に，重力波イベントのうち，あきらかに信号がない部分について，背景ノイズが定常ホワイトガウスノイズであるかの統計的検定を行った．今回は代表例としてGW190412 HanfordとGW191215_223052 Livingston(文献[5]で，定常ホワイトガウスノイズと比べると，KS検定・AD検定のp値が小さいと判定されたイベント)についての結果を示している．GW190412が合成ガウスノイズでの同じ検定と比べてp値が大きく異ならなかったのに対し，GW191215_223052では，p値が全体的に小さくなった．この背景の「荒れ」は，500Hz付近で顕著に見られた．また，検定の前に行うホワイトニングの設定を変えた(fdurationを延ばした)ところ，全体的にp値が大きくなったため，これはO3 Livingstonの510Hz付近に見られる線ノイズ「violin mode」が十分にホワイトニングで除去できなかったためではないかと考えられた．

In the first half, we set up “DINGO-IS,” a machine learning-based parameter estimation software, on a computer equipped with an NVIDIA RTX A5000 and attempted to reproduce the results. As representative examples, we present the results for GW190412 (where the masses of the merging objects differ significantly) and GW190828 (where the masses of the merging objects are similar). The IS efficiency—an indicator of whether DINGO’s inference is performing correctly—was 0.01% for the former and approximately 3% for the latter; the result for the latter, in particular, differs significantly from the results reported in DINGO’s official paper (in the 10% range).
In the second half, drawing on Reference [5], we performed a statistical test to determine whether the background noise in portions of gravitational wave events where no signal was clearly present consisted of stationary white Gaussian noise. In this study, we present results for GW190412 (Hanford) and GW191215_223052 (Livingston) as representative examples (events identified in Reference [5] as having small p-values in the KS and AD tests when compared to stationary white Gaussian noise). While the p-values for GW190412 did not differ significantly from those obtained using synthetic Gaussian noise in the same tests, the p-values for GW191215_223052 were generally smaller. This “roughness” in the background was particularly noticeable around 500 Hz. Furthermore, when the whitening settings were changed (by extending fduration) prior to the tests, the p-values increased overall; therefore, it was considered that this might be due to the “violin mode” line noise observed around 510 Hz in O3 Livingston not being sufficiently removed by whitening.

参考文献:
[1]:“Real-time gravitational-wave science with neural posterior estimation”, Maximilian Dax et al, 2021, https://arxiv.org/abs/2106.12594
[2]:GitHubリポジトリ: https://github.com/dingo-gw/dingo (関連論文へのリンクあり) 2026年9月29日閲覧
[3]:“Neural Importance Sampling for Rapid and Reliable Gravitational-Wave Inference”, Maximilian Dax et al, 2023, https://journals.aps.org/prl/abstract/10.1103/PhysRevLett.130.171403
[4]:“Surge Phenomenon in Optimal Learning Rate and Batch Size Scaling”, Shuaipeng Li et al, 2024, https://proceedings.neurips.cc/paper_files/paper/2024/hash/ef74413c7bf1d915c3e45c72e19a5d32-Abstract-Conference.html
[5]:“Residual Test for the Third Gravitational-Wave Transient Catalog”, Dicong Liang et al, 2025, https://arxiv.org/abs/2509.14924v2
[6]:“Sensitive test of non-Gaussianity in gravitational-wave detector data”, Ronaldas Macas et al, 2023, https://journals.aps.org/prd/abstract/10.1103/PhysRevD.108.063016
[7]:“Does non-stationary noise in LIGO and Virgo affect the estimation of H0?”, Simone Mozzon et al,2021, https://arxiv.org/abs/2110.11731
[8]:“Coalescing Compact Binary Parameter Estimation with Gravitational Waves in the Presence of non-Gaussian Transient Noise”, Yannick Lecoeuche et al,2026, https://arxiv.org/abs/2604.07668v1
https://gwosc.org/O3/o3speclines/

Tadeno SKZN

@tadeno_skzn

スライド一覧

物理学を学んでいます