# Deep learning dark matter map reconstructions from DES SV weak lensing data

**Authors:** Niall Jeffrey, François Lanusse, Ofer Lahav, Jean-Luc Starck  
**arXiv:** [1908.00543](https://arxiv.org/abs/1908.00543)  
**Date:** 2019-08-01  
**Summary page:** https://www.summarizepaper.com/en/paper/1908.00543v1/

## Key points

- First reconstruction of dark matter maps from weak lensing observational data using deep learning
- Trained a convolution neural network (CNN) with a Unet based architecture on over $3.6\times10^5$ simulated data realisations
- Interpret their newly created DES SV map as an approximation of the posterior mean $P(\kappa | \gamma)$ of the convergence given observed shear
- DeepMass method is substantially more accurate than existing mass-mapping methods
- Validation set of 8000 simulated DES SV data realisations showed that DeepMass improved the mean-square-error (MSE) by 11 per cent compared to Wiener filtering with a fixed power spectrum
- N-body simulated MICE mock data also showed that Wiener filtering with the optimal known power spectrum still gives a worse MSE than their generalised method with no input cosmological parameters; improvement driven by non-linear structures in the convergence
- With higher galaxy density in future weak lensing data, deep learning likely to be leading approach for mass mapping with Euclid and LSST
- Study provides evidence that deep learning outperforms traditional mass-mapping methods in terms of accuracy and efficiency

## Summary

This study presents the first reconstruction of dark matter maps from weak lensing observational data using deep learning. The authors trained a convolution neural network (CNN) with a Unet based architecture on over $3.6\times10^5$ simulated data realisations with non-Gaussian shape noise and with cosmological parameters varying over a broad prior distribution. They interpret their newly created DES SV map as an approximation of the posterior mean $P(\kappa | \gamma)$ of the convergence given observed shear. Their DeepMass method is substantially more accurate than existing mass-mapping methods. To validate their approach, the authors used a validation set of 8000 simulated DES SV data realisations and compared their results to Wiener filtering with a fixed power spectrum. The DeepMass method improved the mean-square-error (MSE) by 11 per cent, demonstrating its superior accuracy. Additionally, they used N-body simulated MICE mock data to show that Wiener filtering with the optimal known power spectrum still gives a worse MSE than their generalised method with no input cosmological parameters; they showed that the improvement is driven by the non-linear structures in the convergence. With higher galaxy density in future weak lensing data unveiling more non-linear scales, it is likely that deep learning will be a leading approach for mass mapping with Euclid and LSST. This study provides evidence for this claim by demonstrating that deep learning outperforms traditional mass-mapping methods in terms of accuracy and efficiency. Overall, this research represents an important step forward in our ability to understand dark matter through observations of gravitational lensing effects.