Speaker
Description
Machine Learning Emulators have been widely applied to accelerate cosmological analyses. We develop reliable Transformer architecture models (and some other models) for CMB power spectra in the range of ell=2-5000 within LCDM model, to a precision well bounded by cosmic variance limit. Through improvement in the choice of architecture, activation functions, loss functions, pre-processing of data vectors and sampling methods, we extend both the multipole range and parameter space area of CMB PS emulation. For the freedom of exploring beyond LCDM models in the future, we trained our emulators in a gigantic prior so that we can transfer train our model for extended models and let the users be able to explore the parameter space. We are working on some applications of our models and may be able to demonstrate some results.
