Research

FREmu: Power spectrum emulator for \(f(R)\) gravity (2023 - )

To constrain gravity in the non-linear regime with data from Stage-IV surveys, observables such as non-linear matter power spectra need to be calculated precisely for modified gravity models. However, such observables are often derived from n-body simulations which will cost a lot of resources. In this work, a new public emulator, the FREmu is released which can provide fast and precise predictions of power spectra for the \(f(R)\) gravity model. Our emulator uses artificial neural networks to construct a map from parameters to power spectra and the training data are derived from the Quijote-MG simulation suite. The FREmu has a parameter space of 7 dimensions including \(\Omega_m\), \(\Omega_b\), \(h\), \(n_s\), \(\sigma_8\), \(M_{\nu}\), \(f_{R_0}\) and its accuracy is better than 5% for most cases.

Test modified DGP model with cosmological data (2023 - 2024).

In cosmology, the DGP gravity model provides a screening effect for the formation of large-scale structures but fails to comply with cosmological observations in the absence of dark energy. In this work, we investigate cosmology under the DGP gravity model, focusing on dark energy models within this gravitational framework to reconcile with observational cosmology. We consider two forms of dark energy, namely brane dark energy and bulk dark energy. To constrain the models, we utilize a range of observational datasets including cosmic chronometers (CC), Type Ia supernovae (SNIa), baryon acoustic oscillations (BAO), and cosmic microwave background radiation (CMB). We perform a joint analysis of these datasets using the Markov Chain Monte Carlo (MCMC) method to obtain the dark energy equation of state that best fits the data, and analyze potential sources of dark energy within the DGP gravity framework.

Accelerate cosmological n-body simulations with deep learning (2024 - )

N-body simulations are playing an essential role in modern cosmology. However, precise simulations will cost lots of resources, so how to make simulations more efficient is a very interesting question. We are now using deep-learning techniques to accelerate n-body simulations without losing any accuracy.