Speaker
Description
In this work, we adopt KNe data to prepare a training, test, and validation data set to be fed into a conditional variational autoencoder to regenerate the KNe light curves for the required values of physical parameters. For different KNe models, the physical parameters governing the light curves are different based on the pre-merger or post-merger properties of the BNS merger event. We train the CVAE on the KNe data by conditioning the light curves on the physical parameters, with a training time of ~20 minutes, and rapidly generate light curves for the desired parameter values. Once the CVAE is trained and conditioned on the physical parameters, it takes ~1 milli-second to generate the light curves with a root mean square value of ~0.02 (AB mag) between the true and generated light curves, thus speeding up the process by ~1000 times as compared to the existing method. We have separately trained, generated, and verified the CVAE approach on two different KNe models, where one model is based on pre-merger while the other is on post-merger properties of BNS, and have obtained satisfactorily accurate results, with training time and light curves generating time of ~20 minutes and ~1 millisecond respectively while achieving a root mean square value of ~0.02 and 0.015 AB mag between the original and generated light curves for each model. This technique has the ability to provide an alternative to the time-consuming and resource-draining simulations.
| Submitted on behalf of a Collaboration? | No |
|---|
