DANSur_HM: Modularly incorporating higher modes in a deep learning based gravitational-wave surrogate
Abstract
Numerical relativity (NR) simulations provide the most faithful representation of the gravitational waves (GWs) emitted by binary black hole (BBH) systems during merger. In the context of GW astronomy, tasks such as parameter estimation can require vast numbers of waveform evaluations per second across the entire parameter space. Since performing full NR simulations for each evaluation is not computationally feasible, interpolating methods for existing NR waveforms, known as surrogate models, have been developed with marked success. In this paper, we build on our previous work to introduce methods to train a fast surrogate model based on neural networks in order to generate BBH merger waveforms, including the fundamental (2,2) mode, as well as the (3,3), (2,1), (4,4), (3,2), (4,3) and (5,5) higher-order modes. Applying a pretraining step on approximant data before fine-tuning on NR data allows us to smooth out the parameter space, and making use of the parallelization ability of GPUs to project the NR waveforms in the inclination-phase $(ι, ϕ)$ sphere during training allows the fitting of all the explored modes simultaneously. The developed surrogate model, \texttt{DANSur\_HM}, achieves average mismatches of the order of 10−4, with the worst mismatch at 2.5×10−3, and achieves throughput above 6×105 waveforms/second on an NVIDIA V100 GPU. Parameter estimation tests confirm the usefulness of the inclusion of higher modes.
Related
- Gravitational-wave parameter estimation with machine-learning generated surrogate waveforms
- Architecture Acceleration of Machine Learning Gravitational Waveform Surrogate Models
- A Pairwise Surrogate for Gravitational-Wave Spectra from Highly Relativistic Vacuum Bubble Collisions
- Deep Learning Detection of Beyond-General-Relativity Deviations in Gravitational-Wave Signals: A Detection-Threshold Study with Real LIGO Noise
- Polar coordinate transformations for machine-learning-based dark matter subhalo detection in strong gravitational lenses
BibTeX
@misc{freitas2026dansurhm,
title = {{DANSur_HM: Modularly incorporating higher modes in a deep learning based gravitational-wave surrogate}},
author = {Osvaldo Gramaxo Freitas and Anastasios Theodoropoulos and Nino Villanueva and José A. Font and Antonio Onofre and Alejandro Torres-Forné and José D. Martin-Guerrero},
howpublished = {arXiv},
year = {2026},
month = sep,
doi = {10.48550/arxiv.2609.03025},
eprint = {2609.03025},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.03025},
}
RIS
TY - UNPB
TI - DANSur_HM: Modularly incorporating higher modes in a deep learning based gravitational-wave surrogate
AU - Osvaldo Gramaxo Freitas
AU - Anastasios Theodoropoulos
AU - Nino Villanueva
AU - José A. Font
AU - Antonio Onofre
AU - Alejandro Torres-Forné
AU - José D. Martin-Guerrero
PY - 2026
DA - 2026/09/15
PB - arXiv
DO - 10.48550/arxiv.2609.03025
UR - https://arxiv.org/abs/2609.03025
AB - Numerical relativity (NR) simulations provide the most faithful representation of the gravitational waves (GWs) emitted by binary black hole (BBH) systems during merger. In the context of GW astronomy, tasks such as parameter estimation can require vast numbers of waveform evaluations per second across the entire parameter space. Since performing full NR simulations for each evaluation is not computationally feasible, interpolating methods for existing NR waveforms, known as surrogate models, have been developed with marked success. In this paper, we build on our previous work to introduce methods to train a fast surrogate model based on neural networks in order to generate BBH merger waveforms, including the fundamental (2,2) mode, as well as the (3,3), (2,1), (4,4), (3,2), (4,3) and (5,5) higher-order modes. Applying a pretraining step on approximant data before fine-tuning on NR data allows us to smooth out the parameter space, and making use of the parallelization ability of GPUs to project the NR waveforms in the inclination-phase $(ι, ϕ)$ sphere during training allows the fitting of all the explored modes simultaneously. The developed surrogate model, \texttt{DANSur\_HM}, achieves average mismatches of the order of 10^−4, with the worst mismatch at 2.5×10^−3, and achieves throughput above 6×10^5 waveforms/second on an NVIDIA V100 GPU. Parameter estimation tests confirm the usefulness of the inclusion of higher modes.
ER -