Einstein-Net: Modelling Black Holes using Error-Bounded Neural Networks
In this paper, we investigate the use of neural networks for generating black-hole initial data that satisfy the constraint equations of general relativity. Existing neural approaches to solve this problem often enforce the constraints via loss penalties without providing error estimates for individual predictions. It is noteworthy that any solver in this domain must compute new initial data for each choice of mass, momentum, and spin. Here we introduce Einstein-Net, in which the Bowen–York momentum constraint is satisfied exactly, and the puncture structure, positivity, asymptotic flatness and Schwarzschild limit are incorporated into the output formula. Thus, in our case, only the Hamiltonian constraint remains to be learnt. To do this, we represent the Arnowitt–Deser–Misner mass explicitly using the Gauss-law identity. We also derive an analytical error bound from the Hamiltonian residual, evaluate it numerically, and assess a separate acceptance rule using independent reference cases. Our experiments on single black holes with parallel momentum and spin provide median mass errors of a few parts in ten thousand. Furthermore, the generated data reproduce the expected apparent horizons and physical mass and spin relations. Thus, we show that the neural network setup we present here can generate black-hole initial data accurately and effectively.
Authors
- Hassan Ugail (ORCID: https://orcid.org/0000-0002-3084-1797)
Institutions
- University of Bradford (GB)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-19
- DOI
- https://doi.org/10.21203/rs.3.rs-11776989/v1
- Primary Topic
- Pulsars and Gravitational Waves Research
- Type
- preprint