WenhanGuo/MIU2Net

Weak lensing mapping of dark matter based on ML methods

MIU²Net

Weak Lensing Mass Inversion

MIU²Net stands for Mass Inversion U²Net. It uses deep learning to convert weak lensing shear ($\gamma$) maps to convergence ($\kappa$) maps, which trace the projected dark matter distribution.

We develop MIU²Net as a deep learning framework for weak lensing mass inversion. MIU²Net includes observations effects like shape noise, reduced shear, and data masks in the training.

Installation

The main MIU²Net package depends on the following packages:

• pytorch
• numpy
• scipy
• astropy

Prior to installing MIU²Net, make sure to install PyTorch according to your OS and compute platform. We recommend installing pytorch 1.12.0 and torchvision 0.13.0 to avoid unexpected dependency errors. The main MIU²Net package includes the full training and testing code for deep learning.

To reconstruct convergence maps using traditional (non- deep learning) methods, we have modified the cosmostat package developed at the CosmoStat Lab in CEA Paris-Saclay, so that we can use traditional $\kappa$ map reconstructions alongside deep learning for comparison, or even use traditional methods during network training. Avaiilable reconstruction methods from cosmostat include:

• Kaiser-Squires (KS) deconvolution
• Wiener Filtering (WF)
• sparse reconstruction ((Lanusse et al. 2016), Glimpse)
• MCALens Starck et al.

To use these methods within MIU²Net, you should install Sparse2D developed by the CosmoStat Lab. This is not required by the deep learning framework.

MIU²Net train/test

To train a MIU²Net model:

cd ./miu2net/main
python train.py xxxxxxxxxx

Testing MIU²Net model:

cd ./miu2net/main
python pred.py xxxxxxxxxx