Prepare a plan for dissemination and outreach activities for the entire lifetime of the Action, with regular updates each 12 months.

Setup a github repository with python code example and documentation, developed in the WGs, using Machine Learning solutions.

One of the goals of the Action is to uncover the most suitable Machine Learning techniques for detecting and analysing gravitational waves and for managing detector noise and seismic disturbances. While we work on different research directions to achieve this, we collect both python code we used and experimented with as well as multimedia documentation on the Machine Learning techniques of interest. This will contribute to the dissemination of our concrete Machine Learning findings to the scientific community at large and will provide a tried-and-tested collection of tools for our own future collaborations.

The collection is available at the github repository:
https://github.com/orgs/g2net-CA17137/repositories

Reports on possible HPC splution based on Deep Learning strategy for Gravitational Wave glitches classification.

Some selected publications from the ones reported by our G2net Action, https://www.g2net.eu/publications-and-talks, address this deliverable:

(1) Scalable auto-encoders for gravitational waves detection from time series data.
Roberto Corizzo, Michelangelo Ceci, Eftim Zdravevski, Nathalie Japkowicz.
Expert Systems with Applications 151, 113378 (2020). DOI: 10.1016/j.eswa.2020.113378

(2) Core-Collapse supernova gravitational-wave search and deep learning classification.
Alberto Iess, Elena Cuoco, Filip Morawski, Jade Powell.
Machine Learning: Science and Technology 1, 025014 (2020). arxiv:2001.00279 DOI: 10.1088/2632-2153/ab7d31

(3) Predicting the Properties of Black-Hole Merger Remnants with Deep Neural Networks.
Haegel, S. Husa.
Classical and Quantum Gravity 37, 13 (2020). arXiv:1911.01496. DOI: 10.1088/1361-6382/ab905c

Report on suitable ML pattern-recognition techniques for identification of non-stationary spectral lines and non-stationary noise sources.

Some selected publications address this deliverable:

(1) Machine-Learning Nonstationary Noise Out of Gravitational-Wave Detectors.
Gabriele Vajente, Yiwen Huang, Maximiliano Isi, Jenne C. Driggers, Jeffrey S. Kissel, Marek J. Szczepanczyk, Salvatore Vitale.
Physical Review D 101, 042003 (2020). arXiv:1911.09083. DOI: 10.1103/PhysRevD.101.042003

(2) Application of dictionary learning to denoise LIGO’s blip noise transients.
Alejandro Torres-Forné, Elena Cuoco, José A. Font, Antonio Marquina.
Physical Review D 102, 023011 (2020). arXiv:2002.11668. DOI: 10.1103/PhysRevD.102.023011

Report on suitable machine learning techniques and solution for the search of Gravitational Wave signals from pulsars.

Some selected publications address this deliverable:

(1) BinarySkyHough: A new method to search for continuous gravitational waves from unknown neutron stars in binary systems.
B. Covas, Alicia M. Sintes.
Physical Review D 99, 124019 (2019). arXiv:1904.04873. DOI: 10.1103/PhysRevD.99.124019

(2) The Adaptive Transient Hough method for long-duration gravitational wave transients.
Miquel Oliver, David Keitel, Alicia M. Sintes.
Physical Review D 99, 104067 (2019). arXiv:1901.01820. DOI: 10.1103/PhysRevD.99.104067

(3) Generalized application of the Viterbi algorithm to searches for continuous gravitational-wave signals.
Joe Bayley, Chris Messenger, Graham Woan.
Physical Review D 100, 023006 (2019). arXiv:1903.12614. DOI: 10.1103/PhysRevD.100.023006

(4) Deep-learning continuous gravitational waves.
Christoph Dreissigacker, Rahul Sharma, Chris Messenger, Ruining Zhao, Reinhard Prix.
Physical Review D 100, 044009 (2019). arXiv:1904.13291. DOI: 10.1103/PhysRevD.100.044009

(5) First all-sky search for continuous gravitational-wave signals from unknown neutron stars in binary systems using Advanced LIGO data.
B. Covas, Alicia M. Sintes.
Physical Review Letters 124, 191102 (2020). arXiv:2001.08411. DOI: 10.1103/PhysRevLett.124.191102

Report on the monitoring of the seismic noise for GW detector and on its impact and impact on Newtonian Noise.

The analysis of low-frequency gravitational waves (GWs) requires multidisciplinary research in the fields of seismic sensing, signal processing, robotics, machine learning and mathematical modeling

So far, we have:

(1) Initiated the interdisciplinary community and identified a common research ground.

(2) Drafted a general framework for multisensory fusion of mobile spatially distributed sensors for low frequency geographic dynamical signals which goes beyond the state of the art.

(3) Performed initial tests on off-the shelf irobot create 2 robot which have confirmed the viability of the approach and suggested the development of ad-hoc platforms (the seismic sensors work well on a create 2 in the bandwidth 1-10 hz).

(4) Triggered a cooperation between Heron Robots (team led by F. Bonsignorio) and Astrocent (team led by T. Bulik) that is developing a physical ad hoc platform for seismic noise sensing. It is worth notice that this cooperation was triggered by the G2 meeting in Rijeka. The level of advancement can be estimated at 80% at the moment of the submission of this report. The platform will be made available to G2net and Virgo members in the appropriate forms.

(5) A master thesis summarizing the state of the art has in parallel be completed: https://www.ge.infn.it/~chinca/INFN/Pulze_thesis_complete.pdf

Report on Design study for application for robots to adaptively monitor seismic noise around GW detectors.

A detailed design study for application for robots to adaptively monitor seismic noise around GW detectors was performed as a part of g2net network and stakeholder activities. So far, we have:

(1) Initiated the interdisciplinary community and identified a common research ground. In particular, we triggered a cooperation between a team led by Prof. F. Bonsignorio, Heron Robots (http://www.heronrobots.com/) and a team led by  Prof. T. Bulik, University of Warsaw (https://indico.ego-gw.it/event/464/contributions/4137/attachments/2271/4013/MobileSeismicCOST28-09-2022.pdf) that is developing a physical ad hoc platform for seismic noise sensing, a mobile robot. An ‘alfa’ prototype has been already built and tested. New prototype iterations are under development. The cooperation represents a prominent outcome of the g2net networking activities and it was triggered by the 3rd g2net MC Meeting, which was held at University of Rijeka (Croatia): https://indico.riteh.hr/event/1/ .

(2) Performed initial tests on off-the shelf irobot create 2 robots which have confirmed the viability of the approach and suggested the development of ad-hoc platforms: https://www.g2net.eu/wp-content/uploads/2023/05/technical_report.pdf

(3) Drafted an operational roadmap identifying the actions to be performed and the technical/scientific gaps to be overcome to achieve the objective of significant Newtonian noise suppression and low-frequency noise characterization. A general framework for multisensory fusion of mobile spatially distributed sensors which goes beyond the state of the art is proposed: https://www.g2net.eu/wp-content/uploads/2023/05/roadmap_bonsignorio.pdf

Report on Design study for application for robots to adaptively monitor seismic noise around GW detectors.

Report on new ML method application in seismological problems for GW detection and related experiments.

Some selected publications address this deliverable:

(1) Rapid Prediction of Earthquake Ground Shaking Intensity Using Raw Waveform Data and a Convolutional Neural Network.
https://doi.org/10.1093/gji/ggaa233
This study describes a deep convolutional neural network (CNN) based technique to predict intensity measurements (IMs) of the earthquake ground shaking at stations far from the epicentre which have not yet recorded the maximum ground shaking. The CNN IM predictions use raw waveform data and do not require previous knowledge of the earthquake source (location and magnitude) Although the technique is not strictly designed for earthquake early warning, we find that it can provide useful estimates of ground motions within 10-15 s after the earthquake origin time depending on various setup elements (e.g. times for data transmission, computation, latencies).

(2) Transfer learning: improving neural network based prediction of earthquake ground shaking for an area with insufficient training data.
https://doi.org/10.1093/gji/ggab488
In the previous study, we showed that convolutional neural networks (CNNs) applied to network seismic traces can rapidly predict earthquake peak ground motion intensity measures (IMs) at distant stations using only recordings from stations near stations the epicentre. We used 10 s, raw, multistation (39 stations) waveforms for the 2016 earthquake sequence in central Italy for 915 M ≥ 3.0 events, the CI data set, which has a large number of spatially concentrated earthquakes and a dense network of stations. In this work, we applied the same CNN model to an area of central western Italy with a smaller-sized data set, getting worse performance than the results presented in the previously published study. To counter the lack of data, we successfully explored the adoption of ‘transfer learning’ (TL) methodologies and show that the use of TL improves the results in terms of outliers, bias and variability of the residuals between predicted and true IM values. We demonstrated the ground shaking warning times that could be received for the seismic station ‘IV.PII’, located 10 km from VIRGO.

(3) Intra-domain and cross-domain transfer learning for time series data— How transferable are the features?.
https://doi.org/10.1016/j.knosys.2021.107976
This research explores the transferability of features across different domains of time series data and assesses the impact of transfer learning on predictive power and convergence rate during training. Four different machine learning architectures were utilized for the experiment, each trained on the small datasets with and without transfer learning. The efficiency of transfer learning was evaluated within the same application domain of seismology, as well as across different domains including speech, medicine, and finance. The study assessed the compatibility between different source and target domains for knowledge transfer and analyzed the impact of the size of the target dataset, model choice, and hyperparameters on transfer learning. Results indicate that transfer learning is very likely to either improve or not negatively impact a model’s predictive performance or convergence rate Importantly, we expect our findings to be applicable to any time-series domain, where small dataset sizes appear as a constraint for machine learning applications.

(4) The Choice of Time–Frequency Representations of Non-Stationary Signals Affects Machine Learning Model Accuracy: A Case Study on Earthquake Detection from LEN-DB Data.
https://doi.org/10.3390/math10060965
This study aimed to explore the possibility of alternative time-frequency representations offering superior information for machine learning classification of seismic signals. Specifically, we employed three prominent convolutional neural networks to classify seismic waveforms into either earthquake or noise categories, utilizing nine distinct time-frequency representations. Our findings demonstrate that the Pseudo Wigner-Ville and Wigner-Ville time-frequency representations outperform the benchmark model significantly, whereas the spectrogram, widely used in non-stationary signal analysis, does not produce statistically significant gains in performance. Furthermore, our results suggest that different time-frequency representations could be employed in diverse fields that use non-stationary time series data to extract valuable information when working with machine learning analysis of those data (e.g. geophysics, voice recognition, EEG and ECG signals, gravitational waves, etc.).

Report on the application of ML methods in the control systems for GW detectors.

Working Group 3 (WG3) explored the use of Machine Learning (ML) techniques in the control and noise mitigation strategies of scientific experiments, specifically for Gravitational Wave (GW) detectors. GW detectors, both those currently running and those foreseen to be spaceborne, are uniquely complex instruments with specific and new challenges in terms of control and noise issues. These challenges call for significant adaptation and ingenuity in the ML approaches, which are seldom used as textbook cases and are often coupled with simulations and burden with heavy experimental constraints. We have worked towards developing ML algorithms as part of the detectors’ feedback-control systems as well as for the feed-forward cancellation of noise.

The WG3 training school taking place in Fall 2021 held two more related lectures, one on ML specifics and one on using fractal analysis in controlling interferometers:
· “Deep Learning Taxonomy”, presented by Razvan Pascanu (DeepMind, UK).
· “Fractal analysis for interferometer control”, presented by Marco Cavaglia.

The link to the presentation materials is here:
https://indico.ego-gw.it/event/217/timetable/#all.

Regarding the cancelling of Newtonian Noise, Alessio Cirone and Andrea Chincarini used Neural Networks instead of the traditional method of Wiener Filter. This resulted in Alessio Cirone’s PhD thesis on the use of deep neural networks as an alternative filter to cancel Newtonian Noise in current and future Gravitational Wave (GW) detectors. The link to the thesis is here: https://iris.unige.it/handle/11567/1001554.

In an online WG3 workshop in Spring 2021, Diego Bersanetti discussed “Controls in Advanced Virgo: from Lock Acquisition to Lock Losses” and led a round table discussion on how Machine Learning can contribute to this area. A link to the presentation materials is here: https://indicoego-gw.it/event/172/timetable/#20210322.detailed.

In the same event, Fiodor Sorrentino addressed “Squeezed light benches and optical alignment issues”, with a follow-up discussion on the Machine Learning contribution in this topic, link here: https://indicoego-gw.it/event/172/timetable/#20210323.detailed.

Two more relevant lectures took place in this event, namely one on non-stationary noise removal in LIGO data and another on glitch removal in GW detectors:
· “Non stationary noise removal from LIGO data”, presented by Gabriele Vajente.
· “Glitch removal in ground-based gravitational-wave interferometric detectors”, presented by Marco Cavaglia.

The link to the presentation materials for these two lectures can be found here: https://indico.ego-gw.it/event/172/timetable/#all.detailed.

In the WG2-3 meeting in Turku, on June 8, the following talks addressed the application of machine learning to GW detectors:
· “Towards a neural network based sensing & control for GW observatories”, presented by Nikhil Mukund.
“Combinatorial optimization for sensor placement with deep reinforcement learning”, presented by Conor Muldoon.
· “Newtonian noise subtraction in 2G and 3G detectors using neural networks”, presented by Soumen Koley.

The link to the presentation materials for these papers can be found here: https://indico.ego-gw.it/event/409/timetable/#20220608.

At the fourth and final training school, two lectures were relevant to the control of detectors with ML:
· “Newtonian Noise”, presented by Soumen Koley.
· “ML for control systems”, presented by Nikhil Mukund .

Relevant materials for these presentations can be found here: https://github.com/niksterg/g2net_4th_training_school_thessaloniki_2023.