Skip to main content
SearchLoginLogin or Signup

Evaluating Transposed Convolutional Neural Networks to Generate X-ray Pulsar Model Light Curves

Presentation #301.03 in the session Stellar/Compact II.

Published onMay 03, 2024
Evaluating Transposed Convolutional Neural Networks to Generate X-ray Pulsar Model Light Curves

We present innovative Bayesian Markov Chain Monte Carlo (MCMC) models that employ physical semi-analytic methods and Neural Network (NN) techniques for determining the posterior distributions of the parameters that describe the multipolar magnetic field structures that reproduce the NICER thermal X-ray bolometric light curve of the millisecond pulsar J0030+0451. Our physical semi-analytic approach employs thermal X-ray emission from surface hot-spots that align with the polar cap regions corresponding to multipolar magnetic field structures. While the physical model-based semi-analytic method provides accurate results, it is computationally expensive. In contrast, our NN method (trained on physical models) speeds up the computations by a factor >300 demonstrating an astonishing advancement in efficiency. In order to enhance the robustness of our analysis, we implement various statistical techniques (e.g. Kolmogorov–Smirnov test, Wasserstein distance, Kullback–Leibler divergence, and Jensen–Shannon divergence) to quantify comparisons between the posterior distributions derived from the semi-analytic and NN approaches. Furthermore, we will discuss the optimal NN performance for different loss functions and training dataset sizes. Finally, we will discuss how this broader approach can be applied to various waveforms from other sources in different wavelengths and messengers.

Comments
0
comment
No comments here