A research agent for blazars and a neural network for the photon ring

Astrophysics today is a logistical discipline. It has to pull together a fast-growing literature, heterogeneous data from across the spectrum, and physical models that keep getting more complex. Some teams are trying to organize that work rather than automate one task at a time.

A multi-agent system for blazars

Authors from Armenia, Italy and the United States describe the AstroGenesis framework, among them David Pasham and Alessandro Tramacere. It is a domain-specific multi-agent system. It combines literature search and synthesis, access to multi-frequency data and its analysis, theoretical modeling, and the generation of research directions. A supervising agent works with a planner and replanner architecture. At the center sits a theoretical modeling agent that draws on previously trained neural surrogate models. Complex broadband, multi-aspect modeling can then run as a conversation in natural language, without launching full computations every time.

The authors measured the literature search layer above all, on two test sets: one where each question concerns a single publication, and one where the answer requires evidence from several papers. At least one relevant publication showed up in the first five results for 76.6 percent of the questions in the first set and 79.2 percent in the second. Example workflows show how the literature, observational data, modeling and hypothesis generation add up to a traceable analysis. The creators say they plan to expand to other fields of astrophysics.

A ring that reveals spin

A separate paper reaches for the measurement tool itself. Very long baseline interferometry could directly image structures of a black hole on the scale of the event horizon, among them the photon ring. The n = 1 subring is especially valuable, because its radial profile and azimuthal brightness modulation encode information about the black hole's spin. In observations it overlaps with the emission of the n = 0 subring, so it has to be isolated first.

Courtney Duong, Frank Myhre and Joseph Farah trained a convolutional neural network on 111 000 simulated images of rings, generated with the eht-imaging tool. The network reconstructs the radial and angular profile of the n = 1 subring with a mean normalized cross-correlation of about 0.99. It reports no false detections when there is no ring in the data, and the extracted features make it possible to recover the brightness modulation sensitive to spin. On images from the KerrBAM simulation and magnetohydrodynamic ones it reconstructs the general radial profile, though discrepancies appear in the intensity profile. The authors tie this to future missions, such as the planned Black Hole Explorer.