Cosmological Parameter Estimation Under Constrained Simulation Budget with Optimal Transport-based Data Augmentation

Our solution to the Challenge is to generate a large set of cheap lognormal simulations and then use an optimal-transport–based mapping to match their distribution to that of the expensive, high-fidelity simulations. We use these OT-corrected maps to pretrain a compact, parameter-efficient EfficientNet model that directly predicts Omega_m​ and S8​ with calibrated uncertainties, and then fine-tune it on the limited set of true simulations. I will show how this lognormal+OT augmentation and pretraining pipeline effectively “stretches” the available simulations and yields competitive performance with a lightweight, simulation-efficient inference model.