Speaker
Description
Determining high-fidelity, flexible parameterizations of hadronic parton distribution functions from experimental data remains a long-standing goal in precision phenomenology. Traditionally, this problem is addressed through dedicated, large-scale global analyses spanning hundreds of experimental datasets and conditioned on underlying theory choices such as perturbative order, potential beyond Standard Model contributions, and many other fitting assumptions. This has produced an ecosystem of PDF determinations that are connected through the underlying QCD theory, but difficult to interpolate between directly. It is within this ecosystem that we present MORPH, a foundation-model framework for learning reusable, uncertainty-aware representations of hadron structure, designed to provide a universal backbone for PDF inference across theory choices and observable modalities. MORPH currently uses self-supervised representation learning to encode ensembles of fitted PDF replicas into a structured latent space that captures the full flavor and x-dependence with calibrated uncertainty reconstruction. We benchmark this representation using a suite of machine-learning and physics-based scaling studies to determine whether it preserves the QCD information carried by the original ensembles. A central aim of MORPH is to condition this latent space on theory metadata inputs, enabling interpolation across PDF phenomenological extractions as a fast-evaluation complement to dedicated global fits. I will present progress in building this foundation-model for PDF ensembles, discuss the physics diagnostics needed to validate such representations, and outline a path toward fast, generative inverse mapping of the non-perturbative structure of the proton relevant for the next generation of particle physics experiments.