2026 workshop program
A General ML Framework for Diagnosing Physics Tensions in the Multi-Probe Era
The work
Abstract
We introduce Classifier Based Tension Analysis (CBTA), a general, model-agnostic machine learning framework for diagnosing the origin of persistent tensions across multiple physics probes. The method frames tension assessment as a likelihood free classifier test between real observational data and forward modeled mock realizations drawn from a joint generative model of shared physics and probe specific systematics. A machine learning classifier is trained to distinguish real from mock summary statistics, with the area under the ROC curve (AUC) serving as a calibrated test statistic: an AUC near 0.5 indicates consistency with known systematics, while a persistently elevated AUC, even under generous nuisance flexibility, signals the need for new physics. The framework is fully general and scalable to arbitrary datasets and probe combinations, encompassing tensions in cosmology, astrophysics, and high energy physics, including H0 and Sigma 8, spectral index, and astroparticle discrepancies. Statistical calibration is achieved through permutation tests and posterior predictive checks. Beyond a binary verdict, CBTA produces ranked ablation tables identifying which systematics or physics extensions drive or resolve a given tension, offering theorists and experimentalists a principled, reproducible roadmap for model building and resource allocation in the multi-probe era.