Monday, July 27th, 2026
Workshop: Mustapha Ishak-Boushaki
Time: 10:00 AM - 11:00 AM
Location: SCGP 102
Title: Cosmic acceleration and dark energy in light of DESI and other recent cosmological observations - Part 1
Speaker: Mustapha Ishak-Boushaki
Title: Cosmic acceleration and dark energy in light of DESI and other recent cosmological observations - Part 1
Speaker: Mustapha Ishak-Boushaki
Tuesday, July 28th, 2026
Workshop: Xingang Chen
Time: 10:00 AM - 11:00 AM
Location: SCGP 102
Title: Exotic Dark Matter- Part 1
Speaker: Xingang Chen
Title: Exotic Dark Matter- Part 1
Speaker: Xingang Chen
Workshop: Xingang Chen
Time: 11:30 AM - 12:00 PM
Location: SCGP 102
Title: Exotic Dark Matter- Part 2
Speaker: Xingang Chen
Title: Exotic Dark Matter- Part 2
Speaker: Xingang Chen
YITP Event: Bela Arwen MA Thesis Defense
Time: 2:00 PM - 4:00 PM
Location: C. N. Yang Institute of Theoretical Physics Common Room 6-125
Title: Transfer Learning for Cosmological Emulators
Abstract: Cosmological parameter inference from Stage-IV surveys such as the Roman Space Telescope requires millions of theory predictions per analysis, each expensive enough that the direct calculation is prohibitive. Neural network emulators solve this by learning the map from cosmological parameters to the predicted measurement, but they move the cost rather than removing it: training one emulator requires tens of thousands of the same expensive calculations, and standard practice starts over from scratch whenever the analysis configuration changes. This thesis asks whether a trained emulator can instead be reused. Transfer learning is tested across two kinds of change: the prescription used for nonlinear structure growth, and an enlargement of the cosmological parameter space itself, from a cosmological constant to an evolving dark energy equation of state. In both cases, fine-tuning a pretrained emulator reaches a given accuracy on substantially less training data than training a fresh one, with the advantage largest where data are scarcest. The conventional economy of freezing layers is shown not to help, while correcting how the pretrained weights are scaled at initialization is shown to matter more than any choice of strategy. Stage-IV pipelines can therefore amortize emulator training across configurations rather than paying for it repeatedly.
Title: Transfer Learning for Cosmological Emulators
Abstract: Cosmological parameter inference from Stage-IV surveys such as the Roman Space Telescope requires millions of theory predictions per analysis, each expensive enough that the direct calculation is prohibitive. Neural network emulators solve this by learning the map from cosmological parameters to the predicted measurement, but they move the cost rather than removing it: training one emulator requires tens of thousands of the same expensive calculations, and standard practice starts over from scratch whenever the analysis configuration changes. This thesis asks whether a trained emulator can instead be reused. Transfer learning is tested across two kinds of change: the prescription used for nonlinear structure growth, and an enlargement of the cosmological parameter space itself, from a cosmological constant to an evolving dark energy equation of state. In both cases, fine-tuning a pretrained emulator reaches a given accuracy on substantially less training data than training a fresh one, with the advantage largest where data are scarcest. The conventional economy of freezing layers is shown not to help, while correcting how the pretrained weights are scaled at initialization is shown to matter more than any choice of strategy. Stage-IV pipelines can therefore amortize emulator training across configurations rather than paying for it repeatedly.