Machine Learning and Galaxy Formation

Galaxy formation and evolution involves the interplay of a large number of complex, non-linear processes, many of which act at scales beneath those accessible to even the most modern galaxy formation simulations. Galaxy formation models therefore include parameterised sub-grid processes, which must be calibrated against selected observational constraints. In this thesis, I explore the application of machine learning and optimisation methods to characterize and calibrate a semi-analytic model of galaxy formation, GALFORM. I investigate the application of deep learning to this problem, building an accurate emulator of the full model over a ten dimensional parameter space from just 1000 GALFORM evaluations. I investigate the calibration of GALFORM to a large number of datasets, and investigate tensions between different choices of calibration datasets and the parameters themselves. Next, I present an investigation into the controversial requirement for a top-heavy stellar initial mass function in starbursts in the GALFORM model, which it was argued was necessary for the model to match the constraints from the number counts of sub-millimeter galaxies, their redshift distribution, and the local K-band luminosity function. Here, I apply Bayesian Optimisation to search the model parameter space for optimal fits to these datasets, and demonstrate that GALFORM is not capable of reproducing these data simultaneously with a solar neighbourhood IMF, and that the top-heavy IMF alleviates this problem.