Cosmology

Finding Features in Cosmology Mass Maps

This project is a Research work done at NERSC as a Summer Research intern with Debbie Bard. My work involved applying Deep Learning methods to identify structures in the Universe. This data was a simulation of Cosmology Mass Maps. During this summer, I developed two Deep Learning models, one to reconstruct important structures in the mass maps and the other to classify which Cosmology model the mass maps come from.
The two techniques developed were Unsupervised Denoising Convolutional Autoencoders and Supervised CNN. I was responsible for coding the model, testing on NERSC's supercomputing nodes and optimizing the model parameters. A picture of me and Debbie with Supercomputer Cori can be found here.
We acheived some good results and later I presented a Poster at 229th American Astronomical Society meeting.

You can see the original image (left) and recontructed features (right) in the image above. This is the result obtained from the Unsupervised Convolutional Autoencoder. The reconstruced image is interesting because we are now able to identify high-concentration areas in the heat map which are nothing but galaxies.

Above is a T-SNE plot which is obatined by representing features abtained from the CNN in a 2D plot. You can see a separation between the features from first Cosmology model (blue) and the second Cosmology model (red).

At 229th AAS meeting