Abstract
Gravitational waves from binary neutron star mergers provide insights into dense matter physics and strong-field gravity, but waveform modeling remains computationally challenging. We develop a deep generative model for gravitational waveforms from BNS mergers, covering the late inspiral, merger, and ringdown, incorporating precession and tidal effects. Using the conditional autoencoder, our model efficiently generates waveforms with high fidelity across a broad parameter space, including component masses (m₁,m₂), spin components (S₁ₓ,S₁ᵧ,S₁z,S₂ₓ,S₂ᵧ,S₂z), and tidal deformability (Λ1,Λ2). Trained on 3×10⁵ waveforms from the IMRPhenomXP\_NRTidalv2 waveform model, it achieves an average overlap accuracy of 99.6% on the test set. The model significantly accelerates waveform generation. For a single sample, it requires 0.12 seconds (s), compared to 0.38 s for IMRPhenomXP\_NRTidalv2 and 0.62 s for IMRPhenomPv2\_NRTidal, making it approximately 3 to 5 times faster. When generating 10³ waveforms, the network completes the task in 0.86 s, while traditional waveform approximation methods take over 46-53 s. Our model achieves a total time of 7.48 s to generate 10⁴ such waveforms, making it about 60 to 65 times faster than traditional waveform approximation methods. This speed advantage enables rapid parameter estimation and real-time gravitational wave searches. With higher precision, it will support low-latency detection and broader applications in multi-messenger astrophysics.