Can physics-informed neural networks evolve black-hole punctures?

A network that solves the vacuum Einstein equations

The paper asks whether physics-informed neural networks (PINNs) can handle the moving-puncture evolution used in numerical relativity (NR). The network takes the two-dimensional coordinate (t, r) as input and outputs four spherically symmetric fields: the lapse α, the radial shift βr, the Z4c conformal factor χ, and the radial conformal metric component γ̃rr. These are exactly the variables that stay regular at the puncture, unlike the physical metric, which blows up as χ→0 near the singularity.

Training minimizes a loss that combines a data term, the Ricci tensor of the reconstructed spacetime, and, in the forecast runs, the moving-puncture gauge conditions. The Ricci residual enforces the vacuum Einstein equations Rμν = 0, while the 1+log slicing and Gamma-driver shift residuals remove the coordinate ambiguity that the covariant vacuum equations alone leave open. The architecture is a separable PINN: two modified MLP sub-networks, one in t and one in r, each with four hidden layers of width 128, combined into a rank-K = 64 representation of the four fields. Exponential parametrization keeps α, χ, and γ̃rr positive, which the authors report was necessary for stable training.

Interpolation reproduces numerical-relativity data

The first test trains the network across the whole NR domain, t ∈ [0, 100] and r ∈ [0.505, 170], using 3.4 × 10⁶ spacetime coordinates on 201 uniformly spaced time slices. Alternate radial grid points at each time slice are held out for testing. With data loss and physics loss enforced everywhere and gauge weights set to zero, the PINN reproduces the rapid early gauge evolution from a pre-collapsed lapse toward the trumpet solution, including the strong-field region inside the horizon where the physical metric becomes very large.

The paper also checks whether the training loss can stand in for the true error when no ground truth is available. Comparing the L∞ error of each predicted field against the nominal loss for that field, the two measures are strongly correlated down to losses of order 10⁻³ to 10⁻². Below that threshold, further loss reduction improves satisfaction of the Einstein equations but not agreement with the NR data, because the NR data itself carries numerical error and does not exactly satisfy Einstein's equations. The authors flag this tension between data loss and physics loss as a recurring theme across their experiments.

Forecasting the puncture from early-time and boundary data

The second test turns interpolation into a Cauchy problem. The network now sees labeled data only for t ≤ 10 (initial data) and r ≥ 150 (outer boundary), while the physics and gauge conditions are enforced over the full domain up to t = 60. The PINN carries the puncture through the transition to the trumpet solution and reproduces the qualitative evolution: the lapse continues to collapse at the puncture and then settles, the shift grows and peaks near the horizon, and the peak of γ̃rr grows and moves outward before settling to the stationary trumpet profile. For α, βr, and χ the PINN and NR curves are nearly indistinguishable.

The clearest discrepancy appears in γ̃rr around r ∼ 20 at later times. The authors attribute it to a small, sharp gauge wavefront propagating outward through the unlabeled region. The Ricci residual is insensitive to coordinate changes and therefore does not constrain this feature, which is controlled only by the gauge conditions; at the residual level the network attains, the gauge loss cannot capture it. Quantitatively, at t = 60 the PINN Ricci residual is of order 10⁻⁵ for the (t, r) components, three to four orders of magnitude above the NR data, whose own components computed with the same finite-difference operator show order 10⁻⁹ error. The authors describe this gap as the finite accuracy reached by the network and its optimization, a known limitation of PINN training, while noting the solution still matches NR qualitatively and quantitatively.

Reconstructing the Schwarzschild interior from exterior data

The third test is an inverse problem. Analytical Kerr–Schild data for a unit-mass Schwarzschild black hole is prescribed only for r ≥ 15, and the physics loss is imposed inward to r = 2, the horizon in Kerr–Schild coordinates. No gauge condition is prescribed, so the problem is ill posed and admits infinitely many solutions. Because the data is noise-free and analytically generated, training and validation mean squared errors reach values of order 10⁻¹⁶, far below the noisy-data case.

The reconstructed interior disagrees with the analytic Kerr–Schild solution by up to 11% (in βr at r = 2), because the PINN converges to a solution in a different gauge. To verify that it still found the expected Schwarzschild spacetime, the authors train a second neural network to learn the coordinate transformation between the PINN gauge and Kerr–Schild gauge. After that transformation, all four fields agree with ground truth to better than 8 × 10⁻⁴ relative down to the horizon. The three time slices lie on top of each other, with fields changing by less than 2 × 10⁻⁴ relative between t = 0 and t = 50, so the PINN recovers the stationary Schwarzschild solution as Birkhoff's theorem requires, in coordinates of its own choosing.

Why the training loss is not the same as the error

Across the three tests, the paper separates two quantities that are easy to conflate. The training objective is available without ground truth; the L∞ error against NR data is not. In the interpolation test the two track each other down to losses of order 10⁻³ to 10⁻², which supports using the loss as a proxy for reconstruction error in that regime. Below that threshold the correspondence breaks, because the loss becomes dominated by derivative terms and the NR reference itself does not exactly satisfy the equations.

The forecast test shows a second distinction. The Ricci residual measures how well the network satisfies the vacuum field equations, and it is insensitive to coordinate changes. The gauge-condition residuals measure whether the solution stays in the same gauge as the NR data. A small gauge wavefront can be nearly invisible to the Ricci residual while still producing the largest visible deviation in γ̃rr. The paper reports that the 1+log loss sits two to four orders of magnitude larger than the NR data error across the entire radial domain, consistent with the network not resolving the smaller gauge features.

What it takes to train these networks

Reaching the reported accuracy required more than a standard PINN setup. The authors combine curriculum training that progressively extends the physics region into the unlabeled domain, a higher density of collocation points near the puncture (half the collocation points are drawn in r ≤ 15, an empirical choice because fields vary steeply there), and a damped Gauss–Newton Levenberg–Marquardt refinement after the first-order AdamW stage. The LM stage uses 1024 labeled points and 2048 collocation points drawn from a scrambled Sobol sequence, held fixed for 15 steps and then redrawn, with 1500 LM iterations per application.

Training ran on a single NVIDIA A100-40GB GPU in double precision and took approximately 8 hours per training. The authors state plainly that PINNs do not yet reach the accuracy of NR, and that reaching a given accuracy costs more than a standard finite-difference or spectral code. They also note that the current work is limited to spherical symmetry and a single non-spinning black hole, and outline extension to binary black-hole spacetimes as future work.

Where PINNs add something NR codes do not

The paper's own framing of value is not raw accuracy but a different capability profile. As an interpolant, the network provides a compact, differentiable representation of NR data that can be used for further study, such as identifying apparent horizons or computing geodesics. The inverse-problem result suggests a task that is difficult for standard evolution codes: reconstructing a region of spacetime that is not directly evolved, for example the interior from asymptotic data alone.

The comparison that matters most is therefore between three regimes rather than a single accuracy number. When data covers the whole domain, the PINN interpolates and its loss is a reliable error proxy down to order 10⁻³ to 10⁻². When data covers only early times and the outer boundary, it forecasts the main features but misses a small gauge wavefront, and its Ricci residual stays around 10⁻⁵ against an NR floor near 10⁻⁹. When only exterior data is given and the gauge is left free, it converges to a different gauge but still recovers the correct Schwarzschild solution once the learned coordinate map is applied. The paper's priority is the last two regimes: they define where a PINN does something an NR code does not, and where its accuracy gap is the price of that reach.

Sources used in this answer