Leveraging a highly accurate physics model essential for next-generation cosmological analysis typically involves significant computational demands. We here propose a solution by integrating a machine learning-based emulator into the BORG algorithm for effective sampling of cosmic initial conditions from non-linear cosmic structures.

N-body simulations are key to cosmology but are computationally expensive. ML can speed them up, but can we trust the results? We propose a new method of combining ML and simulations (COmoving Computer Acceleration - COCA) which can catch mistakes made by ML and correct for them.

The strongest tension within the concordance model of cosmology is a mismatch between the Universe's expansion rate inferred from the cosmic microwave background vs the local distance ladder. We construct a Bayesian hierarchical model to infer this rate locally but without requiring supernovae. We still find a significant "Hubble tension", and achieve unprecedented precision by using BORG velocity fields.