A Foundation Model for Lattice QCD: Learning to Understand the Standard Model and Beyond

Descripción del proyecto

Mejores modelos guían con precisión las investigaciones sobre la nueva física

Las simulaciones de cromodinámica cuántica (QCD, por sus siglas en inglés) en el retículo son fundamentales para obtener predicciones teóricas precisas que guíen los experimentos de evaluación de las brechas en el modelo estándar. Sin embargo, estas simulaciones requieren muchos recursos informáticos y carecen de precisión. En el proyecto FoundLatt, financiado por el Consejo Europeo de Investigación, se pretende abordar esta limitación desarrollando el primer modelo de fundamentos de aprendizaje automático para la QCD en el retículo. El modelo combinará elementos de grandes modelos generales con métodos de aprendizaje automático especializados que aceleran tareas específicas en simulaciones de QCD en el retículo. Se prevé que el resultado sea un modelo de «entrenar una vez, usar para siempre» superior a los que actualmente requieren un gran entrenamiento continuo, y por tanto poco práctico. El modelo favorecerá predicciones muy precisas para guiar los experimentos sobre la nueva física.

Objetivo

"Despite describing fundamental physics with spectacular accuracy, the Standard Model of particle physics is known to be incomplete. As experimental programmes probe its boundaries at higher energies and with better precision, it is critical that measurements are faced with fully controlled theory predictions. Lattice simulations of Quantum Chromodynamics (QCD) form a major pillar of this effort. However, even with increasing computing capacity, many lattice QCD calculations will not be possible without new techniques.

FoundLatt aims to solve this challenge by developing the first machine learning (ML) foundation model for lattice QCD. As exemplified by the highly successful ChatGPT, these large, general-purpose models have recently shown great promise in other contexts. Combining methods from these contexts with specialised ML methods that already accelerate specific lattice QCD tasks, I will create a lattice QCD foundation model that performs multiple challenging tasks over a range of physical parameters. The main innovation of this programme is a ""train once, use forever"" methodology, which will involve a centralised investment of effort that supersedes impractical training of non-reusable models.

This objective will be achieved via three complementary work packages (WPs). My team and I will develop and train a foundation model capable of two critical tasks, importance sampling and precise estimation of observables, first for a reduced model of QCD (WP1) and then the full theory of QCD (WP2). We will also develop a method to learn a ""quantum perfect"" action for QCD, then fine-tune the foundation model to accomplish this (WP3). Such an action has zero lattice discretisation artefacts, thereby removing one of the largest sources of systematic uncertainty. The main output of FoundLatt will be an ML model that is shared openly with the lattice QCD community, unlocking state-of-the-art theory predictions for the next generation of new physics searches.
"