Sign in
Ctrl K

Proton structure

Unravelling Proton Structure with Hyperoptimised Machine Learning

image credits: Shutterstock

At energy-frontier facilities such as the Large Hadron Collider (LHC) or the future Electron Ion Collider (EIC), scientists study the laws of Nature in their quest for novel phenomena both within and beyond the Standard Model of particle physics. In this effort, an in-depth understanding of the quark and gluon substructure of protons and heavy nuclei is crucial to address pressing questions from the nature of the Higgs boson to the origin of cosmic neutrinos.

In this project we developed novel machine learning methods to scrutinise the partonic structure of nucleons and heavy nuclei and in this way, we were able to tackle long-standing puzzles in our understanding of the strong interactions, from the origin of the proton spin to the modifications of the gluonic content of heavy nuclei. For this, in this project we combined an extensive experimental dataset and cutting-edge theory calculations within the NNPDF machine learning framework where neural networks parametrise the underlying physical laws. We have this way laid out the groundwork for a first-ever universal analysis of nucleon structure from the simultaneous determination of the momentum and spin distributions of quarks and gluons and their fragmentation into hadrons.

The key achievements of this project have been:

  • The demonstration that gluons are heavily suppressed in the small Bjorken-x region in heavy nuclei.
  • A full-fledged GPU implementation of the NNPDF fitting formalism, which has enabled a speed up by up to a factor 100 as compared to the original CPU-based implementation. This implementation has been key for cornerstone NNPDF analysis, such as the extraction of the strong coupling constant with a precision which is comparable to the current global average.
  • The development of a novel efficient algorithmic strategy to determine the model hyperparameters such as network architectures which define the ML model used for these QCD analysis.
  • A first determination of the spin structure of the proton using high-precision QCD theory calculations, which reveals novel patterns of the proton spin decomposition and provides input for the upcoming EIC operations.

All these results have been published in leading journals of the field and have attracted a lot of attention at international conferences, where they have been presented upon request of the conference organisers.

We are fully satisfied with the outcome of the project, which managed to achieve its original goals and will become a standard ingredient of the NNPDF fitting framework for the years to come. The next steps are to be integrated polarised and unpolarised proton PDFs, nuclear PDFs, and fragmentation functions into a universal QCD framework, which will enable the full exploitation of the data from future facilities such as the EIC and the High-Luminosity LHC.

Concerning the target audience, these would be other researchers working in ML applications to fundamental science where problems such as the unbiased parametrisation of unknown functions with support on very high dimensional spaces is required. All the codes developed in this work are fully open source and can be installed and used thanks to the extensive documentation that was created during this project. The GitHub repository of the NNPDF code is the main resource to be shared in this respect.

Participating organisations

Netherlands eScience Center
Vrije Universiteit Amsterdam
Natural Sciences & Engineering
Natural Sciences & Engineering

Impact

Output

Team

JR
Juan Rojo
Principal investigator
Vrije Universiteit Amsterdam
TR
Tanjona Radonirina Rabemananjara
Postdoctoral researcher
Vrije Universiteit Amsterdam
Alessio Sclocco
Alessio Sclocco
eScience Research Engineer
Netherlands eScience Center
Aron Jansen
Aron Jansen
eScience Research Engineer
Netherlands eScience Center
Gijs van den Oord
Gijs van den Oord
Senior RSE
Netherlands eScience Center
Merijn Verstraaten
Merijn Verstraaten
Rena Bakhshi
Rena Bakhshi
Programme Manager
Netherlands eScience Center
Elena Ranguelova
Elena Ranguelova
Tech Lead
Netherlands eScience Center
Carlos M. R. Rocha
Carlos M. R. Rocha
Lead RSE
Netherlands eScience Center
RS
Roy Stegeman
Postdoctoral reseacher
University of Edinburgh
JC
Juan M Cruz-Martinez
Postdoctoral reseacher
Durham University

Related projects

GAHTIe

High-throughput GPU computing for New Physics searches with electrons in LHCb

Updated 5 months ago
In progress

Related software

NNPDF

NN

Fitting proton substructure to observation data using neural networks

Updated 5 months ago
127 7