A machine-learning classifier for the postmerger remnant of binary neutron stars

This repository contains the datasets and models employed for the work in arXiv:2408.10678.

Datasets

Pandas dataframes created from the numerical-relativity simulations in the CoRe and SACRA databases, together with simulations from Phys. Rev. D 106 (2022) 4, 044026 ans Phys. Rev. D 109 (2024) 12, 123011. For each system, we con-sider data from the simulation with the highest resolution and the largest extraction radius available. We do not include systems with eccentricity or misaligned spins,  and simulations performed with the general-relativistic large eddy simulations (GRLES) subgrid model for turbolent viscosity.

The dataframes include the following information (columns):

'Mtot': total mass

'Mratio': mass ratio as in the database/paper (not the one to use for the classifier)

'LambdaTilde': mass-weighted tidal deformability

'EOS': equation of state employed for the simulation

'ChiEff': effective inspiral spin

'RemnantKey':  None->the simulation collapsed in within simulation time, with collapse time stored in 'CollapseTime'; 'greater' -> the system does not collapse within simulation time, where the simulation time is stored in 'CollapseTime'; 'stable' -> based on the EOS and system total mass (< TOV mass), the remnant formed is a stable neutron star

'CollapseTime': collapse time as explained above

'Mratio_fixed': mass ratio to use for the classifier, defined as m1/m2 for all systems

'dataset_classA.json' -> complete dataset, used for classifier A

'dataset_classBandC.json' -> same dataset, but without points for which the system does not collapse within simulation time, and the simulation time is less than 25 ms  (used for classifiers B and C)

Models

We provide the three classifiers models, saved as *.pkl files, together with their respective scaler to scale the input data before passing them to the classifier  The models and scalers can be read/loaded with joblib as

model = joblib.load('classifierA_model.pkl')
scaler = joblib.load('scaler_classifierA.pkl')

We remind that:

Classifier A distinguishes between prompt collapse to a black hole and formation of a neutron star remnant

Classifier B discriminates between prompt collapse, formation of a hypermassive neutron star that collapses during the simulation, and formation of a neutron star remnant that does not collapse within the simulation time

Classifier C further discerns between short-lived and long-lived hypermassive neutron stars (see paper for details)

Example notebook

The notebook 'Classifier_GW170817_data.ipynb' provides an example of how to employ the classifiers on the GWTC-1 GW170817 data.

Cite

If you make use of this material, also cite the paper arXiv:2408.10678