Black Hole Black Boxes: Numerical Black Hole Metrics via AInstein Neural Networks

This paper extends the AInstein framework—a Physics-Informed Neural Network (PINN) method—from Riemannian manifolds to Lorentzian signature to solve the Einstein field equations. The researchers represent the spacetime manifold as a product of a two-dimensional Penrose domain and a two-sphere. By embedding the sphere globally through its standard stereographic projection, the network learns an ambient metric that naturally respects the manifold's topology. The model is trained using a suite of geometric objectives, including the vacuum Einstein equations, symmetry constraints (motivated by the Birkhoff-Jebsen theorem), and curvature-based invariants such as the Petrov speciality index.

The researchers first validated the architecture by recovering the maximally extended Schwarzschild geometry. By encoding the vacuum Einstein equation and $SO(3)$ symmetry, the network accurately reconstructed the Schwarzschild metric across the Penrose domain. Following this validation, the team generalized the training objective to search for algebraically general Petrov type I solutions. By incorporating a trapped-surface constraint—which identifies regions where light rays are forced inward—the model successfully discovered potentially novel Lorentzian Einstein metrics that possess a genuinely trapped interior, demonstrating the utility of neural networks in exploring non-symmetric vacuum solutions.

Finding explicit solutions to the Einstein field equations is a foundational challenge in general relativity, particularly for stationary black holes. Traditional numerical methods often require complex coordinate grids, gauge fixing, and boundary condition management. This machine learning approach offers a coordinate-invariant, mesh-free alternative that leverages automatic differentiation to enforce geometric consistency. By treating the metric as a differentiable function, this method provides a powerful tool for constructing and classifying new spacetimes that are difficult to access via standard numerical relativity techniques.