2025
SPT-3G D1 T&E
The SPT-3G D1 T&E release features some of the most precise CMB primary power spectrum measurements to date and is incredibly rich in information. The analysis achieves Hubble parameter constraints comparable to Planck, showing no strong deviation from ΛCDM but hinting at differences with DESI data. I contributed heavily to the cosmological analysis, incl. parameter inference and interpretation, and played a central role in shaping the analysis strategy. Tools that I have developed formed core pillars of the work.
2025
The BAO-CMB Tension
The scalar spectral index ns is a key observable for testing inflationary models. This work explores how differences between BAO and CMB data lead to shifts in ns constraints and cautions against interpreting results in the context of inflation until this tension is better understood. I supported the project with expertise in the latest CMB data sets and helped refine parts of the statistical methodology. In turn, I learned a lot about inflation models.
2025
OLÉ: Online Learning Emulator
OLÉ is an emulator for cosmological inference that trains Gaussian Processes on-the-fly during an MCMC analysis to produce predictions with uncertainty estimates. I contributed a new segmented sampling approach to OLÉ that couples to candl likelihoods and exploits the differentiability of the full pipeline to accelerate analyses.
2024
CMB-lite w/ Autodiff
The construction of compressed CMB-lite likelihoods is trivial when gradient-information is available.
I show this by applying a simple numerical prescription to a real data set and validate the success of the reconstruction using a set of simulations.
This implementation of CMB-lite runs within a minute on a personal computer.
2024
candl
I developed candl: a JAX-based python likelihood for CMB experiments. Thanks to JAX's automatic differentiation algorithm, calculating the derivate of candl likelihoods is easy, fast, and robust. This gives easy access to Fisher matrices and facilitates new ways of analysing CMB data.
2022
SPT-3G 2018 TT / TE / EE
I expanded upon the previous SPT-3G release and added temperature band power measurements to the polarisation data. This work featured applying many Planck-style consistency tests to the measurements, showing agreement across frequencies as well as temperature and polarisation.
2022
Early Dark Energy
Does contemporary CMB data contain signatures of early dark energy (EDE)? Though recent ACT data is known to support EDE, this work shows that Planck polarisation data is also of great interest and lends support to a different EDE model with distinct signatures in the power spectrum. The addition of SPT data further increases the preference for EDE over ΛCDM.
2021
Primordial Magnetic Fields
Inhomogeneities sourced by magnetic fields in the early universe may provide an avenue to resolving the Hubble tension. I contributed to an analysis of primordial magnetic fields using Planck, SPT, and ACT data. The focus of this work is on consistency of these datasets, which is explored by discecting individual constraints and studying the expected constraining power using mock data.
2021
SPT-3G 2018 EE / TE
I lead the high-level analysis of the first SPT-3G power spectra. This involved computing the covariance of the band powers, writing and verifying the CosmoMC likelihood, running the MCMC chains, and interpreting the results. I not only looked at the constraints placed on ΛCDM, but also studied extensions to the standard model of cosmology with an eye for the Hubble tension.
2021
Covariance Conditioning
In CMB analysis, covariance matrices are typically conditioning in order to guarantee robust parameter cocnstraints. In this paper, I benchmarked the performance of various commonplace conditioning strategies using a set of simulations, finding that in some cases minimal conditioning can lead to a 30% increase of the parameter uncertainty.
2018
Extreme Digitisation
Antarctica- and space-based CMB experiments rely on satellite links to transmit their data, which are limited by their bandwidth. For my MSc project, I explored a lossy compression scheme for the time-ordered data, which instead of using 32-bit floats, restricts each datapoint to 1-, 2-, or 3-bits. The latter increases the map noise level by <2%.
