ニューラルネットワークによるブラックホールの第1光子リング抽出

Very-long-baseline interferometry (VLBI) may enable direct imaging of fine, event-horizon scale black hole structures such as the photon ring. In particular, the $n=1$ photon subring encodes information about black hole spin in its radial profile and azimuthal brightness modulation, making it a key proxy for spacetime properties. However, the $n=1$ subring overlaps with emission from the $n=0$ subring in observations, requiring isolation of the $n=1$ subring for measurements. We present a convolutional neural network (CNN) capable of extracting the $n=1$ subring from composite images containing $n=0$ and $n=1$ subrings. We simulate $111,000$ $m$-ring images with \texttt{eht-imaging} to train and validate the CNN. The CNN accurately recovers the radial and angular brightness profiles of the $n=1$ subring with a mean normalized cross correlation (NXcorr) of $\sim0.99$. The CNN provides a detection capability by predicting no false positives when input images lack a $n=1$ subring. We perform feature extraction with \texttt{ringfit} to show that the CNN-predicted $n=1$ subring accurately recovers the ground truth radial profiles and azimuthal brightness modulation, demonstrating the CNN's capability to extract spin-sensitive features. We further test the CNN on black hole images generated by \texttt{KerrBAM} and general relativistic magnetohydrodynamic (GRMHD) simulations, finding that it recovers the overall $n=1$ subring radial profiles while exhibiting discrepancies in the recovered intensity profiles. These results demonstrate the potential of deep learning methods to isolate the $n=1$ subring from overlapping emission, providing a framework for analyzing future high resolution black hole images from proposed VLBI missions such as the Black Hole Explorer (BHEX).

12 pages, 10 figures, accepted in ApJ