PREDICTION OF BIANCHI TYPE-I COSMOLOGICAL MODELS USING DEEP LEARNING

Abstract

Accurate modelling of dark matter distributions is essential for estimating cosmological parameters like matter density, amplitude of matter fluctuations, and scalar spectral index. Traditional numerical methods are computationally expensive, while existing deep learning approaches, mainly CNNs, struggle with capturing long-range dependencies in 3D spatial data. This study fills this literature gap by outlining a 3D Swin Transformer-based system that takes advantage of hierarchical feature extraction and self-attention in order to raise prediction precision. On Pycola-simulated dark matter maps, the model reaches a Mean Squared Error (MSE) of 0.0021 and R² value of 0.998, an order of magnitude better than with CNN-based alternatives. These observations illustrate greater efficiency and scalability with a robust system for cosmological parameter estimation for large datasets.