Simulation-based inference of gravitational wave signals with domain-specific optimizations

Fast and reliable inference of gravitational wave source parameters is essential for identifying short-lived electromagnetic counterparts and reducing systematic biases in large event datasets. Amortized simulation-based inference offers a promising solution, delivering near-instantaneous posterior distributions by pretraining on synthetic data. We present enhancements to this method by incorporating domain-specific optimizations, including the relative-binning algorithm to compress data by heterodyning against a reference waveform, combined with a low-dimensional signal representation that efficiently encodes reference waveforms. Additionally, we utilize a tailored coordinate system to simplify the structure of the posterior. Together, these advancements streamline the architecture of the deep learning model, reducing the model complexity and improving generalizability across waveform models and future gravitational wave detector networks.