Please read these instructions before posting any event on Fermilab Indico

The FERMI(FNAL) network authorization method will be removed on Tuesday, Feb 4th at 5PM CST. See news for more details.

16–19 Aug 2021
US/Central timezone

Surrogate modelling of kilonova spectra using conditional variational autoencoders

16 Aug 2021, 09:45
15m

Speaker

Kamile Lukosiute (University of Amsterdam)

Description

In this work, we use variational autoencoders to build a surrogate for the model spectra of optical counterparts to neutron star containing gravitational wave events. Optical counterparts to gravitational wave events reveal information that is not necessarily included in the gravitational wave signal. Modeling of these radioactively-powered time-domain transients, kilonovae, is computationally intensive and yields spectra at discrete times. To use such models within analysis frameworks where a continuous model is required, surrogate models are built, often with further simplifications to reduce data dimensions and thus computation time. Machine learning techniques offer an alternative approach to common surrogate model building methods. We explore using conditional variational autoencoders to build a neural network based surrogate model for kilonova spectra models. We find that, while the surrogate model struggles to reconstruct the spectra exactly, the agreement in bolometric luminosities is within 3%, signaling that the model is learning the general structure of the data. Further, we provide a detailed error verification study on the models. Our model seems to work well enough to be appropriate for use within current kilonova analysis studies.

Primary authors

Kamile Lukosiute (University of Amsterdam) Brian Nord (Fermilab) Zoheyr Doctor

Presentation materials