Conditional Derivative GANs for Parameter-Conditioned Synthesis of Gravitational Waves from Core-Collapse Supernovae

Abstract

Abstract Gravitational wave signals from core-collapse supernovae encode critical astrophysical information, including key astrophysical parameters such as the ratio of rotational kinetic to gravitational potential energy at bounce, pre-collapse differential rotation, and the nuclear equation of state. However, the limited number of observed core-collapse supernovae events makes it challenging to build robust data-driven models. The generation of synthetic gravitational wave signals from core-collapse supernovae can therefore provide a valuable tool for data augmentation in machine learning pipelines, particularly for the training and testing of detection and parameter estimation algorithms. Although recent efforts have explored the use of generative models to synthesize waveform-like data, existing approaches lack the ability to condition outputs on specific astrophysical parameters. To address this, we propose a model capable of generating gravitational wave signals conditioned on astrophysical parameters of core-collapse supernovae rotation. The proposed conditional derivative generative adversarial network is trained on the Richers et al. catalog, which contains 1,824 simulated time-series gravitational wave waveforms and their corresponding physical parameters. Quantitative evaluation using time-domain similarity and distributional metrics shows that the generated signals reproduce the key statistical and morphological features of the original dataset. This work introduces a conditional generative framework for rotating core-collapse supernovae gravitational wave signals, that enables fast, low-cost data augmentation and parameterized modeling, improving ML-based detection and parameter estimation.