Developing solvents for unclogging the calculational bottleneck in high-energy physics

Searching for the building blocks of the Universe, trying to answer fundamental questions such as, "What is Dark Matter, really?", particle physicists today have to perform computationally very expensive calculations on supercomputers in order to compare theory with experiment. This project aims to remove existing computational bottlenecks that currently hinder us from making new discoveries. This will be achieved in a cross-disciplinary collaboration between particle physicists and statisticians by developing advanced statistical methods for searches for new elementary particles, creating new types of machine learning that can reduce the computational cost in such searches, and developing more modern code for more efficient calculations. During the project's first two years the focus has been on two main tasks. The first task has been to build code for continual learning on physics data. This is a machine learning technique that constantly updates itself on a stream of data. Here, a codebase building on a statistical method called Gaussian process regression is near completion, as are two accompanying papers describing the code and applying it to practical problems. These are planned published in the fall of 2024. The second task has been to build a robust codebase for computing so-called production cross sections in Quantum Field Theory. This describes the probability of producing new particles in a particle accelerator such as the Large Hadron Collider at CERN, and this calculation is one of the major bottlenecks in particle physics today. We have worked on writing code that is significantly faster than existing code, and which automates calculations required for new physics models. During this work we have identified several inaccuracies and errors in existing code which is heavily used in the field, and found a way to use graphics cards (GPUs) to accelerate numerical calculations.

When exploring the smallest fundamental constituents of the Universe physicists are faced with very serious calculational bottlenecks. To compare new physics models to data, for example from the Large Hadron Collider or astrophysical observations, we need to perform very computationally expensive calculations in quantum field theory (QFT); expensive due to the increasing complexity of higher-order quantum corrections. These are today too slow to perform at the necessary precision except for the simplest models. At the same time, the interpretation of the models given the available data, the best-fit regions of their parameter spaces, and the comparison of different models with each other through their goodness-of-fit, is made computationally intractable due to the size of the parameter spaces of the models and the complexity of the likelihood evaluation for each model. The solution to these inherent problems can not be found in physics alone. This project builds on an interdisciplinary collaboration between physicists and statisticians focused on statistical learning and inference problems in high-energy physics. The project will develop machine learning based regression techniques to speed up QFT calculations with a proper probabilistic interpretation of uncertainties from higher-order contributions, it will develop a continual learning framework for faster emulation of the likelihood for model parameters, and it will investigate new improved statistical approaches to the problems of best-fit and goodness-of-fit using these emulations. Finally, the project will investigate a number of promising new physics models that can answer questions such as: What is dark matter? Are the properties of the Higgs boson indicating that the Universe is fundamentally unstable? Why is there more matter than anti-matter in the Universe? If funded, this project will increase the reach of large and costly experiments in answering these questions for a relatively small extra cost.