Orateur
Description
Parton Distribution Functions (PDFs) are a cornerstone of modern collider phenomenology, encoding the non-perturbative structure of the proton and entering virtually every theoretical prediction at hadron colliders. Their precise determination is essential across broad physics program: from accurate predictions of Standard Model benchmark processes such as Higgs, W and Z boson production, to electroweak precision tests, and the interpretation of potential Beyond the Standard Model signatures. A primary challenge in global PDF fits is achieving rigorous uncertainty quantification when extracting probability distributions from data with complex experimental and theoretical correlations
In this talk, I will discuss the implementation of Bayesian Neural Networks (BNNs) as a framework for PDF determination. Unlike standard neural network approaches that rely on point estimates of network weights, BNNs treat parameters as probability distributions, enabling a statistically principled Bayesian inference in which the posterior over network weights, and consequently the PDFs, is determined by the experimental likelihood given a prior. This provides a natural and direct mapping of data uncertainties onto the final PDF error bands, and offers a rigorous alternative to the widely-used replica method. I will first outline how Artificial Neural Networks are used in PDF determination, then discuss the integration of BNNs into existing fitting frameworks and discuss prospects for disentangling genuine physics signals from model bias and experimental noise, a challenge that sits at the heart of precision QCD.