PhD project: Data science

Project description

This project aims to bring into high-energy physics recent ideas and breakthroughs from Artificial Intelligence (AI). The hope of an easy discovery of a new fundamental theory at the Large Hadron Collider, though well-motivated theoretically, has been thwarted by the first years of data taking, and it is increasingly clear that hints of new physics must be sought for in small deviations from the current theory. This forces us to stretch the current methodology, and inject radically novel ideas in the field. The main bottleneck to discovery at the LHC is knowledge of the structure of the colliding protons, specifically as encoded in Parton Distribution Functions (PDFs): the distributions of their constituent quarks and gluons. This project aims at redefining the way PDFs are determined. This redefinition will involve the cross-fertilization between high-energy physics and AI, and developments in the theory and phenomenology of the strong interactions.

An accurate computation has become mandatory in PDF determination. PDF determination generally involves a difficulty of principle: PDFs are extracted from a large, but inevitably discrete set of measurements, which must be used for the determination of several continuous functions (the various quark, antiquark and gluon PDFs). Hence, one is trying to determine several functions from a discrete set of data, which is of course a mathematically ill-posed problem. Moreover, an estimate of the uncertainty on PDFs is also called for. Hence, PDF determination involves the determination of a probability distribution in a space of functions. The NNPDF collaboration has pioneered the study of these problems, using Neural Networks for the parametrization of the PDFs, and Monte Carlo methods to study the probability distributions in the space of functions.

This project is focused on recent breakthroughs in Artificial Intelligence. The recent techniques of deep reinforcement learning and Q-learning have enabled the design of algorithms which, by keeping track (memory) of previous attempts, automatically learn the rules of a game. For PDFs, this means that it is possible to envisage that the methodology itself would be developed and “learned” automatically. For instance, rather than tuning a genetic algorithm (for instance by choosing mutation rates), or choosing a specific “stopping” criterion in order to decide whether the optimum has been found, or optimizing the neural network through the choice of suitable “preprocessing” functions by means of closure test, it will be left to the algorithm to make all these choices, through a suitable learning procedure. This will in turn allow for the use of deep residual networks — neural networks with very high depth (more than 100 layers, to be compared to the two hidden layers currently adopted by NNPDF), and with cross-layer links. The motivation for this is twofold. First, the possibility of creating a single agent — a deep neural network — which learns the whole problem in one sweep (in contrast to the current situation, with one neural network for each PDF). Second, the fact that coupling deep networks with Q-learning — deep Q-learning can lead to very substantial gains in computational efficiency for complex problems. We aim at exploring these avenues, and apply the results to the determination of the next generation of PDF for the LHC in collaboration with Prof. Forte’s team in Milan, who has been awarded an Advanced ERC on this topic this year.

The project includes a secondment with SGI, at the European Research Centre that is being set up in Edinburgh.

Project supervisor

The project supervisor welcomes informal enquiries about this project.

Find out more about this research area

The links below summarise our research in the area(s) relevant to this project:

What next?

More PhD projects