New Generative Models for Parton Distributions

Parton distributions encode the structure of the proton as probed in high-energy collisions and are consequently a crucial ingredient for phenomenology at the LHC. Parton distributions are in principle determined by QCD but we are currently unable to compute them from first principles. They must be therefore be extracted from the data. This is a classic pattern recognition and regression problem, that can be addressed using modern machine learning tools. The NNPDF collaboration has started using AI techniques in this context more than 20 years ago and has pioneered the use of machine learning for accurate uncertainty quantification. The project is in this framework.

Objectives:

The main goal in PDF determination is to achieve a full control of uncertainties. This is a difficult problem because one is determining a probability distribution of probability distributions from a discrete set of datapoints affected by correlated experimental uncertainties, which are related by a convolutional structure further affected by theory uncertainties. The goal of the project is to switch from the current situation, in which  a hyperoptimized machine learning model leads to results whose accuracy is tested a posteriori, to interpretable models in which the relation between experimental, theoretical and modeling uncertainty and the uncertainties in the final outcome can be traced using explainable AI tools.

• To develop probabilistic generative models for parton distributions that go beyond the current perceptron paradigm.
• To design a complete framework that encodes information on the whole probability distribution of the model predictions and not just the expected value.
• To provide faster and reliable tools for model selection of parton distribution functions and its uncertainties.

Expected Results:

The expected results are fourfold:

• An understanding of the way features in the data are encoded in the PDF and conversely.
• An understanding of the way the features are encoded in the machine learning model.
• An understanding of the interplay between theory and model uncertainties
• An understanding of the optimal way modern machine learning models can be used for uncertainty quantification