A Digital Twin for Lattice QCD simulation

Lattice QCD (quantum chromodynamics (QCD)) is a field of study that investigates the dynamics and phenomenology of one of the four fundamental forces of nature, the Strong interaction. Large scale Lattice QCD simulations are carried out in major HPC centres by research groups that span national and international boundaries.

The absence of efficient ways to transfer lattice data between these HPC centres after they have been generated can impede research and in the worst cases result in lost data. Closely related to this is the problem of how and where to store lattice data persisently over long times, so that it can be maximally utilised. Transforming the way data is handled in the lattice community is especially important now as the wider scientific community moves towards embracing the FAIR principles in all aspects of their academic work.

Data generation is the most resource intensive part of traditional non-ML lattice simulations and is therefore a limiting factor that lattice researchers have to consider when planning their research programs. This project is also exploring the application of Machine Learning (ML) techniques to lattice simulations in order to speed up data generation and open up new areas of research.

We are working with many other groups within interTwin to improve the software development practices in the lattice community and to make lattice data storage and retention more efficent, robust, and FAIR-compliant. We are providing feedback to the Data Lake testbed that is currently being tested at DESY, including on how it could aid us in applying the FAIR principles to the lattice data and metadata we need to store. We are also incorporating Software Quality Assurance as a Service (SQAaaS) into our software development workflow to improve the robustness of our lattice code.

We have applied ML methods to simple models on small lattices and are currently working on the extension to more complicated models on larger lattices. We use deep generative models such as normalizing flows to develop alternatives to the standard algorithms used for generating lattice field configurations. The main idea behind the method of normalizing flows is to build and train a neural network to map a theory of interest to another one that is easier to simulate, ideally one in which the degrees of freedom are decoupled. Once such a map is found one can more cheaply draw samples from the original theory of interest