Home

A fully differentiable, end-to-end framework for combined-probes cosmology. The framework supports training JAX-based emulators for rapid cosmological predictions and integrates common cosmology tools and survey data processing to produce theory calculations, covariances, and inference results.

Note: The data used in the pipeline needs to be placed under /combined_probes_data at the top of the directory and is not shipped with this package as it exceeds 2TB. All data is publicly available—please see references within the relevant publications for details on where to find the raw survey data, masks, and foreground maps required. Cosmogrid simulations are available at: http://www.cosmogrid.ai/.

Related Publications

Features

• Fully differentiable framework (JAX-friendly; vmap/jit compatible)
• JAX-based emulator training for rapid cosmological predictions
• Simulation-based covariance matrices (UFalcon/Cosmogrid)
• Flexible YAML configuration for MCMC runs

External dependencies and notes

• The inference runs make use of the UPanda package which is available at this public url: https://cosmo-gitlab.phys.ethz.ch/alexander.reeves/upanda
• When making EDE training data we use AxiCLASS: https://github.com/PoulinV/AxiCLASS. I renamed the Python interface to this axiclassy, so you may see that in some places in the code (for convenience to have two versions fo CLASS available in my environment), but the pipeline works just as well with the standard classy name

Acknowledgments

Developed at ETH Zurich (IPA). Uses esub-epipe for cluster job submission and pipeline orchestration. Thanks to John Hennig and the Euler cluster support team at ETH Zürich for software support.

Feedback

If you have any suggestions or questions about the project, feel free to email me at areeves@phys.ethz.ch. If you encounter any errors or problems, please let me know.