•Structured metadata to describe emulator files. •Also forms configuration file used for input to emulator training and scientific inference frameworks. •Described in publication. doi:10.1093/rasti/rzaf002 •Detailed in public github. https://github.com/HTJense/cosmopower •Versioned release in prep. •In principle can propagate this emulator standard outside of astrophysics! Typical questions of a potential user: •What emulators even exist? •What code and model is this emulator emulating? •Which version? •What range of parameter values is it valid for? •Which observables does it output? •What happens if I try and use it outside these specifications? •Can I extend it? Typical consequences: •Failed MCMC runs (wasted resources). •Erroneous inference from mis-understanding which model is being used. •Confusion when validating results. Why does my result differ to yours? •Wasted effort and resources on re-creating own (and hence trusted) version of emulators. Enhancing Open Cosmology with Emulator Packaging Ian Harrison†, Hidde T. Jense The Science: Cosmology with Angular Power Spectra State of the Art: Computational Speed-up with Emulators Statement of Need: Reliable Emulator Sharing and Re-Use Solution: FAIR-focussed Emulator Packaging Implementation: Lowering Barriers with an Open Framework •Each major science analysis typically involves analysis of ~100 chains which take ~1000 core hours each on HPC clusters. The main computation time is from numerically solving the Einstein-Boltzmann equations. •Traditional code optimisation of Einstein-Boltzmann solvers is reaching limits due to increasing instrumental precision requiring better accuracy and fewer approximations. •Novel physics models involve adding new ingredients (e.g. exotic matter) to the Universe and are even slower to solve for. •Increasingly solved with trained Neural Network Emulators, which learn from pre-generated training data and ~instantaneously predict data from model . •Validation of existing network plus metadata. •Creation of training sets. •Training of networks. •Plotting of accuracy with respect to emulated code. •Interfaces to common inference codes. •Improve pedagogy through making inference possible on a laptop! •Quickly test new data and effects of alternative modelling choices. •Allow others to use new and novel physical models. •Creating and propagating an emulator for your novel model is much easier than propagating the altered numerical code. Data : Intensity fields on the sphere (Cosmic Microwave Background temperature) Model : Evolution of the Universe using Einstein-Boltzmann Equations Comparison space : Angular power spectra Outputs : Contents, history of the Universe Community : Cosmologists and astrophysicists across multiple ~100-person international collaborations Emulators usually created and shared on ad-hoc person-to-person basis! We implemented a user-friendly framework for the end-to-end workflow of NN emulator creation, sharing and use. Model parameters (e.g. fraction of Universe made of dark matter) are inferred using Markov Chain Monte Carlo posterior estimation, comparing models for power spectra to data. We have created standardised “packaging” for Neural Network emulators. †
[email protected] School of Physics and Astronomy, Cardiff University, The Parade, Cardiff, Wales, UK CF24 3AA Priv. comm. 2023? CosmoPower doi:10.1093/mnras/stac064 Universes with different expansion speeds NASA/WMAP Science Team
