Exploring Time-Frequency Representations for Neural Posterior Estimation of Gravitational Waves

Gravitational-wave astronomy demands accurate and efficient inference methods to characterize astrophysical sources from noisy detector data. Traditional approaches rely on likelihood evaluations in either the time or frequency domain, which can be computationally prohibitive. Neural posterior estimation (NPE) offers a flexible simulation-based alternative, but its success depends critically on how the data are represented. In this talk, I explore the use of wavelet-based time–frequency domain representations as a foundation for applying NPE to gravitational-wave signals. Emphasis is placed on comparing different wavelet transforms in terms of their ability to localize signal features across a range of source parameters and noise conditions. Preliminary results highlight trade-offs between resolution, sparsity, and interpretability, suggesting practical guidelines for selecting suitable wavelets for future inference work. This study represents a step toward unifying modern machine-learning inference techniques with physically motivated time–frequency analyses in gravitational-wave astronomy.