a differentiable lensing simulator (preprint, accepted publication)
Weak gravitational lensing is only related to the gravitational field of the foreground matter, and does not depend on the nature, composition and dynamic state of the matter. It is not only the most direct means to study the distribution of dark matter in the universe, but also sensitive to the nature of dark energy, so it is one of the most important probes to study dark matter and dark energy. The weak gravitational lensing signal is mainly represented by a slight distortion (shear) of the galaxy shape. The weak gravitational lensing shear signal can be obtained by measuring the photometric image. The traditional method can extract cosmological information from the weak gravitational lensing shear list through the cosmic shear two-point correlation function, three-point correlation function, peak statistics and other statistical algorithms, and limit the cosmological parameters. Different statistical algorithms can extract limited signals. This project will use deep learning algorithms to maximize the use of the weak gravitational lensing shear signal to optimize the cosmological parameters. I used this code to generate a large number of training samples of the weak lensing mass distribution for the subsequent deep learning cosmological extraction.
MADLens is a python package for producing non-Gaussian cosmic shear maps at arbitrary source redshifts. MADLens is designed to achieve high accuracy while keeping computational costs as low as possible. A MADLens simulation with only 256^3 particles produces convergence maps whose power agree with theoretical lensing power spectra up to scales of L=10000. MADlens is based on a highly parallelizable particle-mesh algorithm and employs a sub-evolution scheme in the lensing projection and a machine-learning inspired sharpening step to achieve these high accuracies.
MADLens is fully differentiable with respect to the initial conditions of the underlying particle-mesh simulations and a number of cosmological parameters. These properties allow MADLens to be used as a forward model in Bayesian inference algorithms that require optimization or derivative-aided sampling. Another use case for MADLens is the production of large, high resolution simulation sets as they are required for training novel deep-learning-based lensing analysis tools.
Download the code, then run
pip install -e .
which installs MADLens together with all its dependencies.
Alternatively, to create a conda environment with the required dependencies, run
conda env create -f MADLens.yml
Run the code with
python run_lightcone.py
Download the code.
For an interactive session, allocate the desired number of nodes, then run
source PrepareInteractiveJob.sh
then run
srun -n <number of tasks> -u python run_lightcone.py <flags>
For job submission, prepare a job script following PrepareInteractiveJob.sh
To display all parameters run
python run_lightcone.py --helpfull
to set a parameter, either change its default in run_lightcone.py or pass them as
python run_lightcone.py --parameter_name parameter_value
Here's a non-exhaustive list of parameters that can be set by the user:
| Parameter | Description | Typical Value(s) |
|---|---|---|
| BoxSize | side length of the simulation box | 128-1024 Mpc/h |
| Nmesh | resolution of the particle-mesh simulation | 64^3-512^3 |
| B | force resolution factor | 2 |
| Nsteps | number of steps in the FastPM simulation | 11-40 |
| N_maps | number of output maps | >1 |
| Nmesh2D | resolution of the convergence map | 256^2-2048^2 |
| BoxSize2D | size of the convergence map in degrees | 2.5-22 degrees |
| zs_source | list of source redshifts | 0.3-2.0 |
| Omega_m | total matter density | 0.32 |
| sigma_8 | amplitude of matter fluctuations | 0.82 |
| PGD | whether to use PGD enhancement or not | True/False |
| interpolation | whether to use the sub-evolution scheme | True/False |
We provide PGD parameters for a limited amount of configurations. Before using PGD, you have to dump those parameters into parameter files by running the notebook PGD_params.ipynb
See lightcone.py for examples of how to compute forward passes, vector-Jacobian and Jacobian-vector products.
For example,
kmaps, tape = model.compute(vout='kmaps', init=dict(rho=rho),return_tape=True)
evolves initial conditions rho into lensing maps kmaps (runs the forward model), while saving operations to the tape.
Once the tape has been created,
vjp = tape.get_vjp()
kmap_vjp = vjp.compute(init=dict(_kmaps=vector), vout='_rho')
computes the vector Jacobian product against a vector of the same shape as kmaps.
To use the MADLens version that support parameter derivatives, checkout the param_derivs branch.
Vanessa Böhm, Yu Feng, Max Lee, Biwei Dai
other packages MADlens relies on: