PINNGraPE: Physics Informed Neural Network for Gravitational wave Parameter Estimation

TL;DR

The work addresses parameter estimation for BBH sources detected with weakly-modelled GW searches, where waveform models are not strongly assumed. It introduces a physics-informed neural network (PINN) framework that embeds the 1.5PN frequency evolution and the mass relation $\mathcal{M} = M_{\mathrm{tot}} \eta^{3/5}$ to infer the mass parameters $M_{\mathrm{tot}}$, $\mathcal{M}$, and $\eta$ from time-frequency representations. A multi-branch CNN processes seven TF resolutions, and its loss combines $df/dt$ residuals at Newtonian and 1.5PN orders with auxiliary algebraic constraints to enforce physics-consistency. The results show relative errors in the mass parameters at the few-percent level and inference times of roughly $2\ \mathrm{ms}$, highlighting the potential for rapid, physics-consistent parameter estimation in weakly-modelled GW searches and guiding future work to include spin effects and direct TF data from cWB.

Abstract

Weakly-modelled searches for gravitational waves are essential for ensuring that all potential sources are accounted for in detection efforts, as they make minimal assumptions regarding source morphology. While these searches primarily target generic transient sources, they are also highly effective at identifying a broad range of compact binary coalescences, demonstrated by the weakly-modelled search algorithm coherent WaveBurst being the first to detect GW150914. Despite their ability to detect compact binaries with diverse properties, the accurate estimation of source parameters from their output remains to be a challenging task. To overcome this, we leverage physics-informed neural networks, which serve as a powerful tool for parameter estimation by applying physical constraints through the universal differential equation governing a compact binary system. With this approach, we rapidly infer the mass parameters of binary black hole merger systems to within 7% from only the time-frequency representation of the gravitational wave signal.

Figures (2)

• Figure 1: Diagram of the multi-branch CNN architecture. 7 TF representations are input into the network, each processed by a different CNN branch. CNN outputs are flattened and inserted into a fully-connected neural network, which directly outputs 3 mass parameters. Physical constraints are considered in the loss terms, used to update the hyper-parameters of the network.
• Figure 2: Relative errors of the inferred mass properties vs. true parameter. The black dashed lines represent zero relative error, while each red marker represents a signal from the test data.