This paper tests whether physics‑informed neural networks (PINNs) can learn the evolution of a black hole from limited information. The authors focus on a simplified case: a single, non‑spinning black hole in spherical symmetry and the commonly used moving‑puncture gauge. They show that a PINN can both reproduce high‑quality numerical relativity data and extend it into regions where no data were given.

The network takes the two coordinates (time and radius) as input and outputs four fields that determine the spacetime: the lapse (which controls how time flows), the radial part of the shift vector (which controls how spatial coordinates move), the conformal factor χ (a rescaled measure of spatial size), and the radial conformal metric component γ_rr (the radial part of the spatial geometry). The training loss combines three terms: a data term that fits whatever numerical or analytical data are provided, a physics term based on the Ricci tensor of the reconstructed spacetime (this enforces Einstein’s equations), and terms enforcing the moving‑puncture gauge conditions. The authors also rescale coordinates and use an exponential parametrization of some outputs to keep fields finite and to aid stable training.

In experiments the PINN acts as a compact, differentiable interpolant of numerical relativity (NR) solutions. When trained on NR data everywhere it reproduces that data. More interestingly, when given only early‑time data and outer boundary data it can evolve the puncture forward in time and recover the main qualitative features of the full NR evolution. That shows the network can learn the dynamics from partial Cauchy data and represent the solution in a smooth, differentiable form.

The authors also study an inverse problem that is mathematically ill‑posed: they give the network only exterior data and no gauge conditions, and ask it to reconstruct the interior. The PINN succeeds in recovering an interior consistent with a Schwarzschild black hole, but it does so in coordinates of its own choosing. The team verifies this by learning the coordinate transformation that maps the learned solution to the standard Kerr–Schild form. This highlights both a strength — the PINN can fill in missing regions — and a caution: without gauge or coordinate fixes the solution is only unique up to coordinate transformations.