Real-Time Gravitational-Wave Parameter Estimation with WebRetriever Neural Networks

Gravitational-wave observations require rapid, accurate parameter estimation to coordinate electromagnetic follow-up of transient events. Standard Markov chain Monte Carlo (MCMC) methods deliver reliable posteriors but take 10-100 hours per event, while neural posterior estimation methods reduce this to seconds yet demand complete retraining at a cost of 10⁴-10⁵ graphics processing unit (GPU)-hours whenever the underlying waveform model changes. We present WebRetriever, a neural architecture organised by renormalization-group principles. Input gravitational-wave signals are separated by perfect-reconstruction filter banks into three frequency bands corresponding to the inspiral, merger, and ringdown regimes. Each band is processed by a dedicated Fourier neural operator encoder with bandwidth-constrained receptive fields, coupled to the other bands via bidirectional cross-scale attention that enforces physical consistency across frequency scales. Evaluated on 90 binary black hole events from the Advanced Laser Interferometer Gravitational-Wave Observatory (LIGO), Virgo, and KAGRA Gravitational-Wave Transient Catalogues, WebRetriever achieves 0.78±0.12~s inference latency with 0.4% chirp mass fractional error, a 10⁵-fold speedup over MCMC, while showing 3-4x smaller accuracy degradation than existing neural methods when tested on waveform families not seen during training. Importance sampling verification confirms posterior reliability at 48% efficiency. For GW170817, the binary neutron star merger with a documented electromagnetic counterpart, WebRetriever recovers chirp mass and sky localisation consistent with standard analyses in 0.9 s. These results demonstrate that scale-decomposed neural architectures generalise across the waveform model landscape, with direct implications for real-time multi-messenger astronomy.