The low luminosity end of Galactic HMXBs with eROSITA: Establishing a luminosity floor for accreting BeXRBs

We present a first look at the Galactic population of heretofore known HMXBs as observed by SRG/eROSITA during its first four surveys. eROSITA’s sensitivity of ∼ 10−13 erg s−1 cm−2, translating to 1032 –1034 erg s−1 in luminosity for most known HMXBs in the Milky Way, has thus far never been reached by any wide-area survey instrument. We present the extended log N-log L distribution of known HMXBs reaching down to 1032 erg s−1 using eROSITA, and show the large scatter that can be induced by source intrinsic variability. We present sub-type resolved luminosity distributions, showing that the Supergiant X-ray binaries (SgXBs) and Be X-ray binaries (BeXRBs) occupy different parts of the overall distribution, and reanalyse RXTE/ASM data and MAXI for comparison to eROSITA. The luminosity regime uncovered by eROSITA allows a systematic study of the “transient” BeXRBs, which are typically below the detection threshold of monitors outside of outburst, and whose low luminosity behavior has been a longstanding question. Signatures of stable accretion at low luminosities have been observed with pointed instruments for a fraction of the overall sample, so far. With the eROSITA results, we posit that accretion outside of outburst is likely the norm, since a vast majority (>80%) of BeXRBs are detected at luminosities at least an order of magnitude higher than expected for the most X-ray luminous Be stars. We discuss the observed luminosity in the context of cold disk accretion and the “propeller” mechanism. We highlight a small subpopulation of “isolated” Be-stars that reach luminosities comparable to the least luminous BeXRBs, hinting at the presence of compact object companions.

Improvements to AI Systems:

1. Astrophysical Source Classification and Subtype Discrimination

• Train a multi-label classifier on eROSITA, RXTE/ASM, and MAXI data to distinguish Supergiant X-ray binaries (SgXBs) from Be X-ray binaries (BeXRBs) based on luminosity distributions and variability patterns.

• The improved AI can automatically classify newly detected X-ray sources into sub-types with confidence scores, enabling real-time triage of survey data.

2. Transient Detection and Outburst Prediction

• Develop a recurrent neural network (e.g., LSTM) that ingests time-series light curves from eROSITA and monitors to predict the onset of BeXRB outbursts, using the newly observed low-luminosity quiescent states as training data.

• The improved AI can forecast transient events hours to days in advance, optimizing follow-up observation scheduling for ground- and space-based telescopes.

3. Luminosity Function Modeling with Intrinsic Variability Correction

• Implement a Bayesian hierarchical model that separates source-intrinsic variability from survey sensitivity limits, using the extended log N–log L distribution down to 10 32 erg/s.

• The improved AI can produce de-biased luminosity functions for HMXBs, correcting for Malmquist bias and variability-induced scatter, leading to more accurate population synthesis predictions.

4. Accretion Regime Identification

• Build a decision-tree or gradient-boosting model that maps observed luminosity and spectral hardness to physical accretion states (e.g., stable cold disk accretion vs. propeller regime), using the >80% detection rate of BeXRBs at quiescent luminosities as a key feature.

• The improved AI can infer the dominant accretion mechanism for individual sources without requiring detailed spectral fitting, enabling rapid physical characterization of large samples.

5. Compact Object Companion Inference for Isolated Be Stars

• Train a generative model (e.g., variational autoencoder) on the luminosity distributions of known BeXRBs and isolated Be stars to identify outliers that likely host undetected compact objects.

• The improved AI can flag candidate binary systems from photometric surveys alone, prioritizing them for radial velocity or timing follow-up to discover new neutron star or black hole companions.

6. Cross-Instrument Data Fusion and Sensitivity Matching

• Create a domain-adaptation framework that aligns eROSITA, RXTE/ASM, and MAXI observations despite differing sensitivities and energy bands, using the overlapping sources as anchors.

• The improved AI can produce a unified, long-term X-ray light curve catalog for all known HMXBs, enabling consistent variability studies across decades and instruments.

7. Simulation-to-Observation Transfer Learning

• Use synthetic HMXB populations (generated from cold disk accretion and propeller models) to pre-train a deep neural network, then fine-tune on eROSITA data to infer physical parameters (e.g., magnetic field strength, accretion rate) from observed luminosity alone.

• The improved AI can estimate key stellar and binary parameters for thousands of sources where detailed follow-up is infeasible, vastly expanding the parameter space for theoretical models.