Gravity bends light. Consequently, adding mass in front of an image will distort the image. Measuring these distortions allow us to determine the distribution of mass that was added. The Rubin Observatory Legacy Survey of Space and Time (LSST) will use this gravitational lensing effect to measure the distribution of dark matter in the Universe, with the goal of better understanding what drives the accelerated expansion of our Universe. Traditionally, this type of analysis is done by calculating summary statistics: think of taking a map of the matter density, and summarizing all that information by, say, the average distance between peaks. This ?compression? is done by necessity: predicting summary statistics is ?easy?, whereas predicting dark matter maps is hard. Scientists at the University of Arizona seek to combine machine learning techniques with recent developments to extract cosmological information from the mass maps themselves. Doing so will ensure that the cosmological information contained in the data will be extracted in an optimal way. As part of this project, the PI will act as a volunteer faculty member in the newly developed TIMESTEP research apprenticeship program at the University of Arizona, training undergraduates on research and coding skills that will improve their ability to secure summer research internships, be it REU programs or in industry. The standard approach for extracting cosmological information from weak lensing survey data relies on the shear correlation function. However, the fact that the matter density field today is non-gaussian renders this approach sub-optimal: if the information contained in its non-gaussian features could be extracted, the cosmological constraints from cosmic shear experiments could be improved by a factor of two or more. Fields-based inference has emerged as the obvious choice for realizing this goal. The team proposes to build on the success of the KARMMA mass mapping algorithm to develop the first practical and accurate field-based inference framework. This work will overcome the need to run dark matter simulations by using approximate methods for modeling non-linear growth. Specifically, the plan is to: 1) extend the KARMMA framework to enable tomographic mass map and cosmological sampling; and 2) train a convolutional neural network to perturb the KARMMA lognormal maps into simulation-quality maps. Efficient sampling of the posterior will be achieved through the use of Hamiltonian Markov Chains. The full inference framework will then be validated using simulated data sets. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.