Gravity-Informed Neural Networks for Post-Newtonian Binary Dynamics
In the authors' words
We introduce a gravity-informed neural-network (GravINNs) framework for learning post-Newtonian binary orbital dynamics. The method incorporates the equations of motion directly into the training phase and is developed at three complementary levels. First, we construct single-orbit surrogates for conservative post-Newtonian dynamics and assess their accuracy across different orbital configurations. We then extend the approach to a parametric model spanning a four-dimensional parameter space defined by the initial separation, radial and tangential momenta, and symmetric mass ratio, allowing a single trained network to represent entire families of orbits and replace repeated numerical integrations. Finally, we consider dissipative dynamics by including the leading radiation-reaction contribution at PN order. In this regime, the network accurately reproduces both the orbital evolution and the associated secular energy loss without using dissipative reference trajectories in the training. These results show that physics-informed neural-networks (PINNs) can provide accurate and flexible representations of both conservative and dissipative post-Newtonian dynamics. The complete implementation and the models used in this work are publicly available through the GravINNs repository.
Appeared: Wednesday, September 23. arXiv. Preprint, not yet peer-reviewed.
Authors' comment: 20 pages, 6 figures. Code available at: https://github.com/JavierMatulich/GravINNs