Flexible Gravitational-Wave Parameter Estimation with Transformers

A single neural network can now adapt to a wide range of analysis configurations when estimating the properties of gravitational-wave sources, according to a new study. The transformer-based model, called Dingo-T1, was tested on 48 events from the third LIGO-Virgo-KAGRA observing run [1]. The work addresses a limitation of existing deep-learning approaches for gravitational-wave inference, which typically require a fixed data-analysis setup. Real observations often demand adjustments such as changes in detector configurations, overall frequency ranges, or localized data cuts to accommodate imperfect signals or to perform specialized tests [1]. The new architecture, detailed in a paper by Annalena Kofler and colleagues, uses a training strategy that allows a single model to handle these variations at inference time without retraining [1]. When applied to parameter estimation, Dingo-T1 analyzed 48 gravitational-wave events from the third LIGO-Virgo-KAGRA Observing Run under a wide range of analysis configurations [1]. The LIGO-Virgo-KAGRA collaboration includes the Kamioka Gravitational Wave Detector in Japan, a 3-km Michelson interferometer built underground and using cryogenic mirrors to reduce thermal and seismic noise [11]. KAGRA participated in the O3 run during 2019 and 2020 and later in O4a for one month in 2023 before returning to commissioning [11]. The model also improved a key performance metric. Median sample efficiency on real events rose from a baseline of 1.4% to 4.2% [1]. Beyond efficiency gains, the researchers demonstrated that Dingo-T1 can enable systematic studies of how detector and frequency configurations impact inferred posteriors, and can perform inspiral-merger-ringdown consistency tests that probe general relativity [1]. The approach provides a principled framework for handling missing or incomplete data, capabilities the authors describe as important for both current detectors and next-generation observatories [1]. The paper was first submitted on 2 December 2025 and revised on 24 June 2026 [1].