Machine learning for lattice QCD
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: High Energy Physics: Lattice. The day in lattice field theory: what the simulations put on the table.
Kai: We're looking at the paper: "Machine learning for lattice QCD".
Mira: Here's the gist of the paper. Machine learning (ML) methods have been explored as a means of optimizing lattice field theory calculations, with increasing interest in recent years This chapter gives an introduction and summary to the use of ML methods to optimize lattice field theory without introducing any uncontrolled approximations The focus is on applications where unbiasedness must be guaranteed, distinguishing them from model-based applications Classification of ML Applications
Paper summary: Kai: So, to summarize this paper "Machine learning for lattice QCD," the central idea is using machine learning methods to optimize calculations in lattice field theory. They are focusing specifically on ways to do this without introducing any uncontrolled approximations into the process.
Mira: And what they claim is that this application of machine learning has been explored over time, and now they are summarizing what's clearly established as being useful right now, marking a point in time that's significant for the study of lattice gauge theory.
Lev: So it’s not just exploring possibilities; it’s looking at specific directions where these methods are already showing promise in practice.
Kai: They highlight techniques like preconditioning and contour deformations as examples of how ML can be used to improve convergence rates, which is a known benefit in numerical solvers.
Mira: But they also emphasize that ML can be leveraged for other things, like learning a mapping between low-precision and high-precision solutions to correct for bias later on.
Lev: That last part about learning a mapping sounds like it could be really powerful if it actually works out when you try to implement it on hardware.
Kai: The paper claims that these ML accelerators can be applied across different steps of the lattice QCD calculation, showing them in a grid of applications.
Mira: So, the significance here is showing how this isn't just theoretical speculation; they are providing a framework for applying these methods where unbiasedness is important.
Lev: It’s about finding ways to handle computationally expensive operations in these simulations in a way that respects the underlying physics structure.
Conclusion: Kai: Looking at "Machine learning for lattice QCD," it seems like the core message is that machine learning is a tool we can use to make these very complex physical simulations more manageable without having to resort to rough approximations.
Mira: That’s right. The authors are pointing toward specific, established methods where ML can give us a computational advantage in lattice field theory calculations while maintaining the rigor of the results.
Lev: From what I hear, it’s about using ML not just to guess answers, but to intelligently navigate the steps of a simulation that are otherwise too slow or too difficult to solve exactly with current methods.
Kai: It’s about finding ways to optimize algorithms by learning how to handle the structure of the data generated by these simulations in a way that speeds things up.
Mira: The implication is that we can use machine learning as an accelerator for fundamental physics calculations, provided we focus on applications where those methods are proven to be unbiased.
Lev: It’s about using ML to tackle the computational wall in areas like quantum field theory, which is a really big deal when you're trying to get real answers about the universe.
Gurtej Kanwar, Scott Lawrence, Alessandro Nada, œnahan
Higgs Centre for Theoretical Physics, The University of Edinburgh · Theoretical Division, Los Alamos National Laboratory · Dipartimento di Fisica, Universitá degli Studi di Torino and INFN · Center for Theoretical Physics – a Leinweber Institute, Massachusetts Institute of Technology · The NSF AI Institute for Artificial Intelligence and Fundamental Interactions
hep-lat, cs.LG
Submitted: 2026-11-06
Updated: 2026-11-06
Comments: 47 pages, 5 figures
Code: https://github.com/FlorianMarquardt/machine-learning-for-physicists
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 68/100
The gist: Machine learning (ML) methods have been explored as a means of optimizing lattice field theory calculations…
Key concepts
- Lattice Field Theory
- This is a mathematical framework used in quantum field theory to study particle physics phenomena like those found in quarks and gluons. Lattice field theory discretizes spacetime into a grid or lattice, allowing for numerical simulations of these complex quantum systems.
- Machine Learning (ML) Optimization
- ML techniques are employed to speed up or improve the calculations within lattice field theory. The goal is to use ML models to guide the simulation process, aiming for better results without making simplifying assumptions that could introduce errors.
- Unbiasedness Guarantee
- This refers to ensuring that the predictions made by the ML optimization method are free from systematic errors or biases introduced by the model itself. This is a key requirement when applying ML to physics problems where accuracy and reliability are paramount.
- Model-Based Applications
- These are applications where a simplified, pre-defined mathematical model is used as the primary tool for prediction. The paper distinguishes its focus from these types of applications, emphasizing methods that do not rely on such models for their core optimization strategy.
Terminology
Summary
Machine learning (ML) methods have been explored as a means of optimizing lattice field theory calculations, with increasing interest in recent years This chapter gives an introduction and summary to the use of ML methods to optimize lattice field theory without introducing any uncontrolled approximations The focus is on applications where unbiasedness must be guaranteed, distinguishing them from model-based applications
Classification of ML Applications
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Surrogate observables can be learned by searching for a neural network that approximates an expensive observable, such as replacing an operator with another function which has
typically small
error without needing to evaluate both on all N configurations. -
Preconditioners can be developed using gauge-equivariant neural networks to enhance numerical properties of linear systems, potentially surpassing traditional heuristics through data-driven optimization and respecting fundamental symmetries inherent in lattice QCD simulations.
-
Generative models, such as normalizing flows, can be trained by minimizing the Kullback–Leibler divergence between the model distribution qtheta and the target Boltzmann distribution p to generate samples from complex target distributions analytically known but hard to sample from.
-
Flow-based importance sampling allows for computing observables by sampling from a known prior distribution q0 and using a transformation f, where the probability density q(U) is explicitly written as q(U) = q0(f−1(U))det J−1.
-
Contour deformations can be parameterized by neural networks to deform the integration contour M on which the integrand is holomorphic, allowing for
deformed observables
Q which have an unchanged expectation value while potentially reducing variance.
Abstract
Machine learning (ML) methods have been explored as a means of optimizing lattice field theory calculations, with increasing interest in recent years. This chapter gives an introduction and summary to the use of ML methods to optimize lattice field theory without introducing any uncontrolled approximations. We focus on those directions which are clearly established as of late 2025, marking the 50th anniversary of the beginning of the study of lattice gauge theory. This chapter is a contribution to the online book ``Lattice QCD at 50 years,'' (LQCD@50), edited by Tanmoy Bhattacharya, Maarten Golterman, Rajan Gupta, Laurent Lellouch, and Steve Sharpe.
Sources
- Adam: A Method for Stochastic Optimization
- A high-bias, low-variance introduction to Machine Learning for physicists
- Applications of Machine Learning to Lattice Quantum Field Theory
- Solving Lattice QCD systems of equations using mixed precision solvers on GPUs
- Equivariant flow-based sampling for lattice gauge theory
- Lattice gauge equivariant convolutional neural networks
- Critical slowing down and error analysis in lattice QCD simulations
- Fighting topological freezing in the two-dimensional CP$^{N-1}$ model
- Large-$N$ $SU(N)$ Yang-Mills theories with milder topological freezing
- Parallel Tempered Metadynamics: Overcoming potential barriers without surfing or tunneling
- Deep Learning Hamiltonian Monte Carlo
- LeapfrogLayers: A Trainable Framework for Effective Topological Sampling
- MLMC: Machine Learning Monte Carlo for Lattice Gauge Theory
- Generalizing Hamiltonian Monte Carlo with Neural Networks
- Accelerate Monte Carlo Simulations with Restricted Boltzmann Machines
- Self-Learning Monte Carlo Method
- Self-Learning Monte Carlo Method in Fermion Systems
- Self-Learning Determinantal Quantum Monte Carlo Method
- Self-Learning Monte Carlo Method: Continuous-Time Algorithm
- Self-learning Monte Carlo with Deep Neural Networks
Related papers
- Heavy-Meson Bag Parameters using Gradient Flow
- Impact of Irrep Truncations on Quantum Simulations of QCD
- Evaluating generative samplers for lattice field theory by scale
- Renormalized symmetry probes for emergent symmetries and non-chiral lattice fermions
- Scalable Generative Sampling and Multilevel Estimation for Lattice Field Theories Near Criticality
- Lattice QCD study of the K*(892) resonance at the physical point