Dynamic Black-hole Emission Tomography with Physics-informed Neural Fields

The Big Picture

Imagine trying to reconstruct a hurricane’s full three-dimensional structure from a single weather station and a handful of blurry snapshots taken minutes apart. That’s roughly the challenge astronomers face when peering into the turbulent gas swirling around a black hole.

The Event Horizon Telescope (EHT) gave us the first photograph of a black hole’s shadow in 2019. But a photograph is flat, frozen, and incomplete. The real action plays out across three dimensions of space and time: superheated glowing gas, hotspots of radiation that flare and vanish in minutes, matter spiraling inward at relativistic speeds.

A team from Caltech, MIT, Princeton, and the University of Toronto has now pushed well past the still image. Their system, PI-DEF (Physics-Informed Dynamic Emission Fields), uses neural networks guided by the laws of physics to reconstruct the full dynamic, three-dimensional gas environment around a black hole from the EHT’s sparse radio data.

PI-DEF simultaneously reconstructs both the glowing gas and its velocity field, enforcing fluid physics as a flexible constraint rather than a rigid rule. This produces far more accurate three-dimensional maps of the black-hole environment than any previous method, along with estimates of the black hole’s spin.

How It Works

The fundamental challenge is a math problem with too many unknowns and too few measurements. The EHT is effectively a single telescope the size of Earth, built by linking radio dishes across multiple continents. It sees the black hole from one angle only, and its data are sparse, noisy, and captured over short time windows. Reconstructing a time-evolving 3D volume from that is what mathematicians call ill-posed: infinitely many 3D scenes could produce the same handful of radio measurements.

The previous state-of-the-art, BH-NeRF, handled this by assuming the gas moves according to Keplerian dynamics, the simple orbital mechanics that govern planets circling the sun. That works reasonably well far from the black hole, where gravity alone dominates. Close in, where extreme gravity and powerful magnetic forces drive the gas, Keplerian orbits are a poor description. Gas can flare up suddenly, vanish, swirl, and splash. BH-NeRF can’t capture any of that.

PI-DEF drops the Keplerian assumption and replaces it with something more flexible:

1. Two neural fields, jointly optimized. The system represents both the emissivity (how brightly the gas glows) and the velocity field (how the gas moves) as coordinate-based neural networks. Each takes a location in space and a moment in time as input and returns a physical quantity like brightness or speed. Both networks train simultaneously.

2. Physics as a soft constraint. Instead of enforcing Keplerian motion as an unbreakable rule, PI-DEF uses the velocity field to guide how emissivity evolves via a continuity equation: the principle that gas doesn’t spontaneously appear or disappear. This enters the learning process as a preference, not an absolute law, so the network can still adapt when turbulence, magnetic activity, or sudden emission departs from smooth fluid behavior.

3. Differentiable rendering through curved spacetime. Light near a black hole bends around the massive object. PI-DEF incorporates gravitational lensing directly into its image-formation calculation, tracing curved light paths from the 3D gas distribution to the simulated EHT measurements. Because the entire pipeline is differentiable, the system can work backward from observed data through the curved-spacetime physics to update its reconstruction.

4. Spin estimation as a bonus. The light-bending calculation depends on the black hole’s spin parameter. By treating spin as an additional learnable quantity, PI-DEF can infer a fundamental property of the black hole from the same data.

On simulated EHT data, PI-DEF clearly outperformed both BH-NeRF and a physics-agnostic baseline. Velocity field recovery improved as well, especially near the black hole where Keplerian assumptions break down most severely.

Why It Matters

The EHT has so far shown us still photographs of two black holes, M87* and Sagittarius A*. Those images confirmed the existence of a shadow, matching a prediction of general relativity. But the real tests of extreme physics require watching the gas move: how matter behaves in ultra-strong gravitational fields, how powerful jets launch outward at near-light speed, how gravity and electromagnetism interact in these environments.

PI-DEF is a step toward making that possible. It also demonstrates a useful design principle. Embedding physical laws directly into a neural network’s training as flexible constraints, rather than applying them as external post-processing, can tame reconstruction problems that would otherwise have far too many solutions. The ability to infer black hole spin from imaging data alone points toward a future where fundamental black hole properties can be read straight from radio telescope arrays.

PI-DEF shows that physics-informed neural fields can reconstruct the dynamic, three-dimensional gas around a black hole far more accurately than methods that either ignore physics or oversimplify it, and may enable direct tests of general relativity in the most extreme environments the universe offers.