Research

My research sits at the intersection of cosmology and machine learning. I work on extracting cosmological information from weak gravitational lensing data, going beyond standard two-point statistics by developing methods that use higher-order and one-point statistics of the cosmic matter and shear fields. A recurring theme in my work is making these analyses more powerful, robust, and interpretable.

On the methodology side, I develop simulation-based inference pipelines that replace analytic likelihoods with forward-modelled simulations. This approach naturally handles non-Gaussian information and complex systematic effects. I have applied this framework in a fully blinded analysis of Dark Energy Survey Year 3 cosmic shear data, and to keep such pipelines physically transparent, I build neural network architectures and attribution methods whose learned features connect directly to N-point correlation functions and cosmic environments. Alongside the machine learning, I work on the analytic side of these statistics — covariance modelling for the integrated 3PCF — and on quantifying systematic effects such as baryonic feedback and intrinsic alignments.

I am a member of the Dark Energy Survey (DES), the Euclid Consortium — where I test and validate SBI compression methods in the forward-modelling group — and the LSST Dark Energy Science Collaboration (DESC), working towards higher-order weak lensing analyses of Stage-IV surveys.

Research Areas

Weak Lensing & Higher-Order Statistics

Using the distortion of galaxy shapes by large-scale structure to constrain cosmological parameters, with a focus on higher-order and one-point statistics beyond the standard two-point functions.

Simulation-Based Inference

Developing likelihood-free inference pipelines that use forward-modelled simulations to constrain cosmology, bypassing the need for analytical likelihood expressions.

Interpretable Machine Learning

Building neural network architectures whose internal representations correspond to known physical quantities, such as N-point correlation functions, to keep machine learning analyses transparent and physically meaningful.

Current Projects

SBi3PCF II: Cosmology from DES Y3 cosmic shear

DES Y3

Applying the SBi3PCF framework to Dark Energy Survey Year 3 data: a fully blinded, joint simulation-based analysis of the cosmic shear 2-point and integrated 3-point correlation functions.

C3NN-SBI on DES Y3

DES Y3

Applying the physics-informed C3NN-SBI compression to DES Y3 cosmic shear, taking interpretable machine learning from simulated fields to survey data.

Interpretable inference with the matter PDF

Methods

Mapping where the one-point matter PDF keeps its cosmological information using neural compression and SHAP attributions, and feeding the attribution maps back into the design of improved summary statistics.

Analytical covariance for the i3PCF

Theory

Deriving the analytical Gaussian covariance of the integrated 3-point correlation function — including patch geometry, pixelisation and estimator-grid effects — and validating it against dedicated Gaussian and N-body mock suites.

Systematics in higher-order statistics

Systematics

Investigating the impact of baryonic feedback and intrinsic alignments on higher-order statistics such as the integrated 3PCF and the one-point lensing PDF.

Euclid DR1 preparation

Euclid

Testing and validating compression methods for simulation-based inference in the Euclid forward-modelling group, preparing the application of higher-order statistics with SBI to Euclid DR1.

Selected Results

Corner plot comparing posterior constraints from the 2PCF alone and the combined 2PCF + i3PCF analysis — Adding the integrated 3-point correlation function (i3PCF) to a standard 2PCF analysis substantially tightens cosmological constraints — validation posterior from the SBi3PCF simulation-based inference pipeline. (Gebauer et al. 2026, JCAP 06(2026)036)

Trained rotationally symmetric convolutional filters of the C3NN architecture — Interpretable machine learning: the trained, rotationally symmetric filters of the C3NN architecture, whose outputs correspond to analytically tractable N-point correlation functions of the input field. (Gong et al. 2024, ApJ 971 156)

SHAP-based signal-to-noise attribution across matter PDF percentiles — Where the matter PDF keeps its cosmological information: SHAP-based signal-to-noise attributions localise the signal in the deep voids and the mildly overdense “shoulder” of the PDF, guiding the design of compressed summary statistics. (Gebauer & Uhlemann, in prep.)

Edgeworth-style buildup of the matter PDF from its cumulants at two smoothing scales — Building the matter PDF from its cumulants: Edgeworth-style expansion and a shifted-lognormal closure compared to the measured distribution at two smoothing scales, connecting the machine-learned summaries back to analytic theory. (Gebauer & Uhlemann, in prep.)