Machine- and deep-learning-driven angular momentum inference from BHEX observations of the $n=1$ photon ring

TL;DR

This paper develops a machine learning framework to infer black hole properties from high-resolution BHEX photon ring images, utilizing a large simulated dataset and novel feature extraction methods, achieving over 90% accuracy.

Contribution

It introduces a new optimized feature extraction method and a minimal set of geometric measures for analyzing the $n=1$ photon ring, enabling effective machine learning-based black hole parameter recovery.

Findings

Achieved over 90% correct spin and inclination recovery.

Developed a feature extraction method outperforming existing techniques by 3000 times.

Validated methods on GRMHD simulations of black hole accretion flows.

Abstract

The $n=1$ photon ring is an important probe of black hole (BH) properties and will be resolved by the Black Hole Explorer (BHEX) for the first time. However, extraction of black hole parameters from observations of the $n=1$ subring is not trivial. Developing this capability can be achieved by building a sample of $n=1$ subring simulations, as well as by performing feature extraction on this high-volume sample to track changes in the geometry, which presents significant computational challenges. Here, we present a framework for the study of $n=1$ photon ring behavior and BH property measurement from BHEX images. We use KerrBAM to generate a grid of $\gtrsim10^6$ images of $n=1$ photon rings spanning the entire space of Kerr BH spins and inclinations. Intensity profiles are extracted from images using a novel feature extraction method developed specifically for BHEX. This novel method is highly optimized and outperforms existing EHT methods by a factor of ${\sim}3000$. Additionally, we propose a novel, minimal set of geometric measurables for characterizing the behavior of the $n=1$ subring geometry. We apply these measurables to our simulation grid and test spin recovery on simulated images using: (i) gradient boosting, a machine learning algorithm; and (ii) an extension of Deep Horizon, a deep learning framework. We find $\gtrsim90$\% correct recovery of BH properties using the machine/deep learning approaches, and characterize the space of resolution-dependent geometric degeneracies. Finally, we test both approaches on GRMHD simulations of black hole accretion flows, and report accurate recovery of spin at the expected inclination of M87*.