Detecting Modeling Bias with Continuous Time Flow Models on Weak Lensing Maps

TL;DR

This paper introduces a machine learning framework using continuous time flow models to detect modeling biases in weak lensing maps, enhancing the validation of cosmological simulations and ensuring robustness against unknown systematics.

Contribution

It demonstrates that field-level probability density estimation with continuous time flow models outperforms other methods in out-of-distribution detection for cosmological data.

Findings

CTFM significantly outperforms feature-level density estimators in OoD detection.

CTFM provides a robust metric for model selection in cosmological simulations.

The approach maintains efficacy across different cosmologies and mitigates inductive biases.

Abstract

Simulation-based inference provides a powerful framework for extracting rich information from nonlinear scales in current and upcoming cosmological surveys, and ensuring its robustness requires stringent validation of forward models. In this work, we recast forward model validation as an out-of-distribution (OoD) detection problem within the framework of machine learning (ML)-based simulation-based inference (SBI). We employ probability density as the metric for OoD detection, and compare various density estimation techniques, demonstrating that field-level probability density estimation via continuous time flow models (CTFM) significantly outperforms feature-level approaches that combine scattering transform (ST) or convolutional neural networks (CNN) with normalizing flows (NFs), as well as NF-based field-level estimators, as quantified by the area under the receiver operating characteristic curve (AUROC). Our analysis shows that CTFM not only excels in detecting OoD samples but also provides a robust metric for model selection. Additionally, we verified CTFM maintains consistent efficacy across different cosmologies while mitigating the inductive biases inherent in NF architectures. Although our proof-of-concept study employs simplified forward modeling and noise settings, our framework establishes a promising pathway for identifying unknown systematics in the cosmology datasets.