Improving cosmological covariance matrices with machine learning

Paper highlights:

• Proposes a convolutional neural network approach to denoise cosmological covariance matrices built from limited numbers of simulations.
• Achieves covariance matrices from only 50–200 halo power spectra that are nearly indistinguishable from those built using thousands of spectra, and generalizes from mock to full N-body simulations.
• Demonstrates that denoised covariances enable cosmological parameter recovery with accuracy comparable to using up to tens of thousands of simulations.

You can read the paper here.