A deep-learning approach that replaces expensive inverse ray-shooting for cosmological microlensing by training a convolutional autoencoder to encode precomputed magnification maps into a compact latent representation and reconstruct them on demand. Students learn how an autoencoder's low-dimensional bottleneck can act as a fast generative surrogate for a physics simulation, and how to design fidelity metrics that decide whether the reconstructed maps are accurate enough to replace the originals for a given observational regime.
Cosmological microlensing occurs when compact objects such as stars in a lensing galaxy further magnify the individual images of a strongly lensed quasar, producing variability in its light curve. Mo…