frb_horndeski_forecast

Open Source Code used in arxiv:2409.11163

Installing

After cloning the repository, create a new conda environment with the provided YAML file

cd frb_horndeski_forecast
conda env create -f env.yml 
conda activate frb_horndeski_forecast

Then, we need to install the provided $\texttt{hiclass}$ package modified by Spurio Mancini et al. 2018:

cd external_packages
tar -xf hi_class_public_modified.tar.gz
cd hi_class_public

make clean
make

Usage

This repository contains three major parts, which are all computationally expensive: Simulation, Training and MCMC sampling.

Simulation

Simulation of (non-)linear matter power spectrum $P_{\mathrm{mm}}^{\mathrm{NL}}(k,z)$, electron bias $b_\mathrm{e}(k,z)$ , ratio of bardeen potentials $\eta(k,z)$, modified gravity change to the Poisson equation $\mu(k,z)$ and comoving distance. Adjust your prefered settings (like prior bounds, number of cores, number of samples) inside the "pk_sim_main.py" file inside the "simulation" directory, then run

python simulation/pk_sim_main.py

The code saves the initial parameter suggestion as "raw_lhc_params.npy", and the successfull simulations and their parameters are stored on the fly in "hiclass_pk_simulation.h5py" and "hiclass_param_dict.h5py", respectively.

The training and testing data of our paper is saved on on OneDrive and can be downloaded here. To train a model with it, download the tar and extract the contents (folders "training" and "validation") into

simulation/output_files/

Training

The $\texttt{cosmopower}$ training settings can be adjusted in the "pk_training_script.py" file. Simply run

python training/pk_training_script.py

to train the model on the data contained in "simulation/output_files/training/". Our pretrained models are stored in "training/trained_models"; feel free to use them.

MCMC-sampling

Run

python mcmc/likelihood_MCMC_NN_main.py

with your preferred settings put into the .py file. The script will look for the cosmopower models inside "training/trained_models/" and use them for the sampling. A minimal working example for calling the models is provided in the notebook "trained_cp_call.ipynb".

References

Feel free to use and adapt this code however you like, but please cite our paper and the corresponding packages you use (cosmopower, hi_class, nautilus, pyccl, hmcode).