KAN-LSTM-Transformer Neural Networks, MFV and Cosmological Parameters

Abstract

Reconstructing the cosmic distance ladder directly from observations is a crucial issue in cosmology. In this paper, we present a novel method for modeling the cosmic distance ladder and estimating cosmological parameters through the use of Kolmogorov-Arnold networks (KAN), Long Short-Term Memory (LSTM), and Transformer networks (collectively referred to as KLT-Net), based on the apparent magnitude data from the Pantheon SN Ia compilation. As a data-driven, non-parametric method for reconstructing the distance modulus $\mu(z)$, KLT-Net is shown to be highly effective in capturing the intricate, nonlinear measurement distributions. After validating against various statistical and machine learning models, we have identified it as the most effective choice among the considered alternatives and ablation experiments. Subsequently, the statistical inference of $H_0$ and $\Omega_{\rm m}$ adopts the flat $\Lambda$CDM framework. Moreover, we introduce the Most Frequent Value (MFV) approach to evaluate the absolute magnitude, $M_B$, from existing literature data. In addition, we employ the Hessian matrix to validate the Bayesian method, demonstrating that the Hubble constant can be precisely constrained from the KLT-Net predictions within the flat $\Lambda$CDM framework. The integration of KLT-Net, the MFV approach, and Bayesian statistics establishes a robust framework for inferring cosmological parameters. This methodology facilitates future cosmological research, particularly in the analysis of complex datasets and the exploration of high-dimensional parameter spaces.