WorksPhysics and Astronomy
Deep learning spacetime from entanglement dynamics
Abstract
In the eternal AdS black hole, the entanglement of the thermofield double state is geometrically realised by the Einstein–Rosen bridge, in line with the ER=EPR conjecture. Hartman–Maldacena surfaces threading this bridge pass behind the event horizon, so the time-dependent entanglement entropy of a pair of half-lines, one on each boundary, probes the black hole interior. We turn this relation around and ask how much of the bulk spacetime can be reconstructed from , using a neural ordinary differential equation that learns the metric directly from the entanglement data. We show that depends on the geometry only through a single function, the local surface gravity expressed in terms of the timelike Killing norm. This is enough to fix a metric with one unknown function: for charged BTZ black holes, the network reconstructs it throughout the region probed by the surfaces, including part of the interior, and the learned metric also yields the Hawking temperature and the location of the critical surface that bounds the accessible interior. With two unknown functions, however, it is not: physically distinct geometries whose two-dimensional sectors are related by diffeomorphisms produce the same . Finally, although the null energy condition is not imposed, the network reproduces data obeying the causal bound on the entanglement velocity, but not data that violate it. Our results thus offer a data-driven perspective on how spacetime, including the black hole interior, emerges from quantum entanglement, and delineate which features of the geometry entanglement alone can determine.