Real-time Classification of Gravitational Wave Events Using Deep Learning
Kowalski, M., Virtanen, J., Järvinen, P., Saarinen, T., & Lehtinen, K.
Astrophysical Journal Letters
DOI: 10.3847/KRIZ.2026.004
Abstract
We present a novel deep convolutional neural network architecture for real-time classification of gravitational wave (GW) events detected by the LIGO/Virgo/KAGRA detector network. Our model achieves 99.7% classification accuracy on a curated test set of 85,000 GW events, reducing the average latency to 1.2 seconds per detection from 12 hours using traditional matched-filtering approaches. The architecture leverages multi-scale temporal convolutions and attention mechanisms to capture complex signal morphologies characteristic of binary neutron star mergers, binary black hole systems, and neutron star-black hole systems. We demonstrate successful generalization across detector networks with varying sensitivities and provide comprehensive uncertainty quantification through Bayesian deep learning extensions. This work enables near-real-time GW classification for multi-messenger astronomy campaigns and rapid electromagnetic follow-up coordination.
Key Findings
- 99.7% classification accuracy on 85,000 validated events
- 1.2 second latency per detection (previously 12 hours)
- Successful cross-network generalization (LIGO, Virgo, KAGRA)
- Bayesian uncertainty estimates for all classifications
KRIZ Data Release 3: 1.2 Million Catalogued Gravitational Wave Events
Virtanen, J., Kowalski, M., Saarinen, T., Metsämaa, R., & Abbott, R.
Astrophysical Journal Supplement Series
DOI: 10.3847/KRIZ.2026.003
Abstract
We present the third major data release from the KRIZ Observatory: a comprehensive catalog of 1.2 million gravitational wave events derived from LIGO/Virgo detector observations spanning 2015-2026. Each event includes signal characteristics, classification labels, sky localization probabilities, and cross-correlations with electromagnetic observations and pulsar timing data. The catalog represents a >40× increase in event count compared to Data Release 2 and incorporates new machine learning classification techniques achieving unprecedented accuracy. Data products include raw strain data, matched-filter results, spectrograms, and pre-computed waveform parameters for 97% of events. All data are publicly accessible via our data portal and API, with cloud storage options on AWS S3 and Google Cloud Platform. This release enables systematic population studies of compact object mergers and unprecedented sensitivity to novel sources.
Dataset Specifications
- 1,247,394 classified events from LIGO/Virgo observations
- 12.4 TB matched-filter data
- 847K time-series datasets (raw strain)
- Complete sky localizations and parameter estimates
- Multi-messenger cross-correlations included
Sub-millisecond Pulsar Detection in Northern Sky Survey Data
Saarinen, T., Lehtinen, K., Kowalski, M., Järvinen, P., & Virtanen, J.
Monthly Notices of the Royal Astronomical Society
DOI: 10.1093/KRIZ.2026.001
Abstract
We report the discovery of 47 new millisecond and sub-millisecond pulsars in archival radio telescope survey data processed through the KRIZ Observatory pipeline. Using advanced signal processing including dispersion measure optimization and spectral whitening across the 400 MHz - 2.5 GHz frequency range, we detected signals with periods ranging from 1.3 ms to 9.8 ms. Follow-up observations with the Nordic telescope network confirmed all candidates as bona-fide pulsars. Seven sources show timing properties consistent with pulsar timing array (PTA) sensitivity to nanohertz gravitational waves. Detailed timing analysis reveals characteristics including spin-down rates, binary companions, and dispersion measures with unprecedented precision. This discovery increases the Nordic PTA sensitivity to gravitational wave backgrounds by 12% and provides targets for continuous gravitational wave searches.
Discovery Summary
- 47 new millisecond pulsars discovered
- 7 sources suitable for PTA observations
- Periods: 1.3 ms - 9.8 ms
- 100% confirmation rate on follow-up observations
- Improved GW sensitivity for Nordic PTA network
Infrastructure Scaling: Processing 4.7 Petabytes per Day on Commodity Hardware
Metsämaa, R., Jokela, T., Kowalski, M., Saarinen, T., & Lehtinen, K.
IEEE Data Engineering Bulletin
DOI: 10.1109/KRIZ.2025.002
Abstract
Large-scale scientific data processing faces significant engineering challenges when throughput demands exceed petabytes per day. We present the architecture and operational experience of the KRIZ Observatory data processing infrastructure, which handles real-time ingestion, processing, and archival of 4.7 PB/day from 12 distributed telescope networks. Our approach leverages standard cloud-grade hardware (NVIDIA A100 GPUs, standard x86 compute nodes) with optimized distributed processing frameworks. Key innovations include: (1) predictive buffer management for heterogeneous data arrival patterns, (2) hierarchical caching strategies across different storage tiers, (3) ML-guided task scheduling for 2,400+ GPU nodes, and (4) automated failover mechanisms achieving 99.99% uptime. We demonstrate cost-effective processing at $0.47 per terabyte and provide detailed guidance for similar large-scale scientific computing systems. Our infrastructure processes data in <2 hours from raw observation to archival, enabling near-real-time science.
Technical Highlights
- 4.7 PB/day sustainable throughput on standard hardware
- <2 hour end-to-end processing latency
- $0.47 per terabyte operational cost
- 99.99% uptime across 3 data centers
- Automated ML-based scheduling for GPU cluster
Nordic Telescope Network: First Light from Kilpisjärvi Station
Järvinen, P., Saarinen, T., Lehtinen, K., Kowalski, M., & Virtanen, J.
Astronomy & Astrophysics
DOI: 10.1051/KRIZ.2025.008
Abstract
We report the successful first light observations from the Kilpisjärvi radio telescope station in Finnish Lapland, the northernmost component of the newly completed Nordic Telescope Network. The station operates at frequencies 327 MHz - 1.4 GHz with 64-element phased array configuration providing 0.3 arcsecond resolution at L-band. Integration with the KRIZ Observatory data pipeline enables real-time data transfer and processing, with science-ready data available within 2 hours of observation. Combined with existing stations in Sweden and Norway, the Nordic network achieves unprecedented angular resolution for northern sky surveys. Early results demonstrate detection of fast radio bursts, pulsar signals, and transient sources with sensitivity exceeding previous facilities by 3-5 magnitudes. This facility represents a significant increase in European observational capacity for radio astronomy and gravitational wave multi-messenger studies.
Facility Specifications
- 327 MHz - 1.4 GHz frequency coverage
- 64-element phased array antenna
- 0.3 arcsec angular resolution
- Real-time KRIZ processing pipeline integration
- Part of Nordic 12-station network
Open Source Tools for Gravitational Wave Data Analysis
Lehtinen, K., Kowalski, M., Virtanen, J., Saarinen, T., & Metsämaa, R.
Journal of Open Source Software
DOI: 10.21105/KRIZ.2025.001
Abstract
We present a comprehensive suite of open-source Python and C++ tools for gravitational wave data analysis, developed and validated through the KRIZ Observatory research program. The toolkit includes: (1) GWCatalog - efficient querying and management of large event catalogs; (2) SignalProcessing - optimized filtering, whitening, and spectral analysis; (3) WaveformModels - fast Python/GPU waveform generation for parameter estimation; (4) SkyLocalization - Bayesian sky position inference; and (5) MultiMessenger - cross-correlation with electromagnetic and neutrino observations. All tools feature comprehensive documentation, extensive test suites, and containerized deployment options. The package has been adopted by 127 research institutions across 34 countries. Performance benchmarks demonstrate 10-100× speedups compared to naive implementations while maintaining numerical accuracy. Source code is available under MIT license with active community contributions and maintenance.
Software Package Components
- GWCatalog - Event catalog management and querying
- SignalProcessing - GPU-accelerated filtering and analysis
- WaveformModels - Fast waveform generation
- SkyLocalization - Bayesian position inference
- MultiMessenger - Cross-correlation and multi-wavelength analysis