J/MNRAS/485/5345 Finding black holes with black boxes (Askar+, 2019)
Finding black holes with black boxes - using machine learning to identify
globular clusters with black hole subsystems.
Askar A., Askar A., Pasquato M., Giersz M.
<Mon. Not. R. Astron. Soc., 485, 5345-5362 (2019)>
=2019MNRAS.485.5345A 2019MNRAS.485.5345A (SIMBAD/NED BibCode)
ADC_Keywords: Black holes ; Clusters, globular ; Models ; Optical
Keywords: methods: data analysis - methods: numerical - methods: statistical -
stars: black holes - globular clusters: general
Abstract:
Machine learning is a powerful technique, becoming increasingly
popular in astrophysics. In this paper, we apply machine learning to
more than a thousand globular cluster (GC) models simulated with the
MOCCA-Survey Database I project in order to correlate present-day
observable properties with the presence of a subsystem of stellar mass
black holes (BHs). The machine learning model is then applied to
available observed parameters for Galactic GCs to identify which of
them that are most likely to be hosting a sizeable number of BHs and
reveal insights into what properties lead to the formation of BH
subsystems. With our machine learning model, we were able to shortlist
18 Galactic GCs that are most likely to contain a BH subsystem. We
show that the clusters shortlisted by the machine learning classifier
include those in which BH candidates have been observed (M22, M10, and
NGC 3201) and that our results line up well with independent
simulations and previous studies that manually compared simulated GC
models with observed properties of Galactic GCs. These results can be
useful for observers searching for elusive stellar mass BH candidates
in GCs and further our understanding of the role BHs play in GC
evolution. In addition, we have released an online tool that allows
one to get predictions from our model after they input observable
properties.
Description:
For the purpose of this study, we used results from numerical
simulations of GC models that were carried out using the MOCCA code
(Hypki & Giersz 2013MNRAS.429.1221H 2013MNRAS.429.1221H; Giersz et al.
2013MNRAS.429.1221H 2013MNRAS.429.1221H) as part of the MOCCA-Survey Database I (Askar et
al. 2017MNRAS.464L..36A 2017MNRAS.464L..36A) project. MOCCA is a code for simulating star
clusters based on Henon's implementation of the Monte Carlo method
(Henon 1971Ap&SS..14..151H 1971Ap&SS..14..151H; Stodolkiewicz 1982AcA....32...63S 1982AcA....32...63S,
1986AcA....36...19S 1986AcA....36...19S) to follow the long-term dynamical evolution of
spherically symmetric stellar clusters.
In this work, we applied machine learning in a new and novel way to
astrophysics: predicting real world properties based on simulations.
An empirical comparison was performed between multiple classifiers
with the guiding metric primarily being a low false positive rate
(F-score). We successfully trained a fairly accurate gradient boosted
decision tree classifier on simulation data and then applied the
learned model on to real world data. K-fold testing with our
simulation data showed that the model had an accuracy of 95 per cent
and a false positive rate of less than 1 per cent. Moreover, applying
the classifier on an independent set of simulations (N-body versus
Monte Carlo) yielded accurate results as well, a classifier trained on
MOCCA Monte Carlo simulations was able to accurately predict the
presence of BHS in N-body (Wang et al. 2016MNRAS.458.1450W 2016MNRAS.458.1450W) simulated
clusters using just the observational properties.
Applied on to real world data, our results are also fairly
encouraging. We ran our model on two different catalogues of Milky Way
GCs in order to get predictions based on independent data sets. We
managed to successfully identify several clusters which have been
ascertained to contain BHs by previous observational studies.
File Summary:
--------------------------------------------------------------------------------
FileName Lrecl Records Explanations
--------------------------------------------------------------------------------
ReadMe 80 . This file
tablea1.dat 65 111 Predictions for the presence of BHSs with data
from the Harris (1996AJ....112.1487H 1996AJ....112.1487H,
Cat. VII/202) catalogue
tablea2.dat 65 84 Predictions for the presence of BHSs with data
from Baumgardt & Hilker (2018MNRAS.478.1520B 2018MNRAS.478.1520B,
Cat. J/MNRAS/478/1520) catalogue for Milky Way
GC parameters
--------------------------------------------------------------------------------
See also:
VII/202 : Globular Clusters in the Milky Way (Harris, 1997)
J/MNRAS/478/1520 : Milky Way globular clusters data (Baumgardt+, 2018)
Byte-by-byte Description of file: tablea1.dat tablea2.dat
--------------------------------------------------------------------------------
Bytes Format Units Label Explanations
--------------------------------------------------------------------------------
1- 9 A9 --- Name Cluster name
11- 15 F5.2 pc R50 Half-light radius
17- 23 E7.2 Lsun/pc2 CSB Central surface brightness
25- 29 F5.2 km/s CVD ? Central velocity dispersion
31 A1 --- f_CVD Flag on CVD (1)
33- 39 E7.2 Lsun TVL Total V-band luminosity
41- 48 F8.2 Myr HMRT Half-mass relaxation time
50- 53 F4.2 pc Rad Core radius
55- 59 A5 --- BHSpred [False/True ] Predictions for the presence
of Black Hole Subsystems (BHS) from the
classifier that was trained on all
MOCCA-Survey I GC models that survive up to
12 Gyr
61- 65 A5 --- BHSfall [False/True ] Predictions for the presence of
Black Hole Subsystems (BHS) from the
classifier that was trained only on
simulated results in which mass fallback
was enabled and BH kicks were lower
--------------------------------------------------------------------------------
Note (1): Flag as follows:
* = Less accurate models, those trained without central velocity dispersion
as a feature
--------------------------------------------------------------------------------
History:
From electronic version of the journal
(End) Ana Fiallos [CDS] 10-Oct-2022