Machine learning approaches in Lattice QCD - an interdisciplinary exchange

10 Jun 2025, 09:00 → 13 Jun 2025, 19:00 Europe/Zurich

HCI G7 (ETH Zürich)

Marina Krstic Marinkovic (ETH Zurich), Roman Gruber

This is the second edition of the workshop previously held at TU Munich in 2023 (https://indico.ph.tum.de/event/7116/).

 

This edition of the workshop will take place at ETH Zurich from 10th to 13th June 2025 and will serve as an interdisciplinary exchange to accelerate the development of machine learning techniques in Lattice Gauge Theories. The workshop will bring together experts in Machine Learning, Lattice QCD, and related fields. The last day of the event will feature international and local experts focusing on the development of efficient simulation software for various GPU architectures, with contributions from both the Lattice QCD and HPC communities.

 

Participants of this workshop are encouraged to also attend the PASC25 conference, which will be held the week following this workshop, from 16th to 18th June 2025, at FHNW in Brugg-Windisch. The PASC25 venue is conveniently accessible by public transportation in less than an hour from Zurich main station.

 

Confirmed speakers include:

• Alexander Rothkopf (Korea University) 
• Alberto Ramos (U. Valencia)
• Andreas Kronfeld (Fermilab)
• Fernando Romero-Lopez (University of Bern)
• Gert Aarts (Swansea University) 
• Steven Gottlieb (Indiana University)
• Tom Luu (FZ Jülich/University of Bonn)
• Will Detmold (MIT)
• Jessica Howard (UCSB)
• Evan Weinberg (NVIDIA)
• Urs Wenger (U. Bern)

 

Local organizers (ETHZ):

• Javad Komijani
• Roman Gruber
• Marina Marinkovic
• Letizia Parato
• Tim Harris

 

Advisory Committee:

• Gert Aarts (Swansea University and ECT*) 
• Nora Brambilla (TUM)
• William Detmold (MIT)
• Lukas Heinrich (TUM)
• Andreas Kronfeld (Fermilab)
• Marina Marinkovic (ETH Zurich)
• Maria Paola Lombardo (Firenze INFN)
• Mike Peardon (Trinity College Dublin)
• Alexander Rothkopf (Korea University)
• Sinéad Ryan (Trinity College Dublin) 
• Phiala Shanahan (MIT)

 

Tuesday 10 June

09:00 → 09:50 Registration

09:50 → 10:00 Welcome

Speaker: Marina Krstic Marinkovic (ETH Zurich)

10:00 → 10:50 Optimal transport for anomaly detection at LHC

Speaker: Jessica N. Howard (Kavli Institute for Theoretical Physics)

Howard_Jessica_ML4LatticeQCD2025_FinalWithBuilds.pdf

10:50 → 11:30 Coffee break

11:30 → 12:20 Machine learning RG-improved SU(3) gauge actions

I describe how RG-improved SU(3) gauge actions can be parametrized through machine learning gauge covariant convolutional neural networks. I discuss how the approach benefits from the straightforward accessibility of gauge field derivatives and the capability to generate targeted learning.

Speaker: Urs Wenger (University of Bern)

FPA_LGE-CNN_wenger.pdf

12:20 → 14:00 Lunch

https://ethz.ch/de/campus/erleben/gastronomie-und-einkaufen/gastronomie/restaurants-und-cafeterias/hoenggerberg.html

14:00 → 14:35 Stochastic differentiation of Monte Carlo simulations for parameter inference in quarkonium suppression

In many scientific domains, phenomena are being described by demanding Monte Carlo simulations. A common problem setting is that these simulations depend on input parameters, whose values are a priori not clear. To determine the input parameters one usually falls back to fitting the output of the simulator to some target. This can become particularly challenging, when the simulator is expensive to evaluate and the number of input parameters increases. In this talk I will discuss the approach of differentiating the simulator, which enables parameter inference by gradient descent. For Monte Carlo simulations involving discrete probabilities it is possible to build on the REINFORCE gradient estimator to differentiate the full stochastic simulation. I demonstrate this method for the simulation of quarkonium suppression. Quarkonium suppression refers to the phenomenon of bound states dissociating in the quark gluon plasma and can be measured by the nuclear modification factor R_{AA}. The underlying simulator that predicts this suppression solves a Lindblad equation by sampling stochastic trajectories. I showcase how to obtain a low variance gradient estimator and fit the quarkonium transport coefficients to nuclear modification factor data.

Speaker: Tom Magorsch

ML4Lattice_2025_magorsch.pdf

14:35 → 15:10 Optimizing static quark states with Neural Network parametrized general Wilson loops

Speaker: Julian Mayer-Steudte

Zurich_ML.pdf

15:10 → 15:40 Coffee break

15:40 → 16:30 Protecting continuum symmetries on the lattice - exact prior information for reinforcement learning

Speaker: Prof. Alexander Rothkopf (Korea University)

ETH_InverseProblem_Discussion_EXPORT.pdf

ETH_InverseProblem_Discussion_FLATTENED.pdf

ETH_Rothkopf_FINAL.pdf

16:30 → 17:15 Discussion

Speaker: Nora brambilla

panelMLEFT2025.pdf

17:30 → 19:00 Welcome Apéro

https://maps.app.goo.gl/gFzm4v1YQ4S4Waj9A

Wednesday 11 June

10:00 → 10:35 Monte Carlo estimates of flow fields for sampling and noise problems

Speaker: Gurtej Kanwar (University of Edinburgh)

Kanwar_ZurichML4Lattice_PDF.pdf

10:35 → 11:20 Coffee break

11:20 → 12:10 Applications of flow models to the generation of correlated lattice QCD ensembles

Machine-learned normalizing flows can be used in the context of lattice quantum field theory to generate statistically correlated ensembles of lattice gauge fields at different action parameters. In this talk, we show examples on how these correlations can be exploited for variance reduction in the computation of observables. Different proof-of-concept applications are presented: continuum limits of gauge theories, the mass dependence of QCD observables, hadronic matrix elements based on the Feynman-Hellmann approach, and the computation of glueball correlators. In all cases, statistical uncertainties are significantly reduced when machine-learned flows are incorporated as compared with the same calculations performed with uncorrelated ensembles or direct reweighting.

Speaker: Fernando Romero López

MLzurich.pdf

12:10 → 13:50 Lunch

13:50 → 14:25 Applications of Stochastic Normalizing Flows to gauge theories and defects

In recent years, flow-based samplers have emerged as a promising alternative to standard sampling methods in lattice field theory. In this talk, I will introduce a class of flow-based samplers known as Stochastic Normalizing Flows (SNFs), which are hybrid algorithms combining neural networks and non-equilibrium Monte Carlo methods. I will then demonstrate that SNFs exhibit excellent scaling with volume in lattice SU(3) gauge theory. Afterward, I will discuss theories with defects and present a general strategy for applying flow-based samplers to such systems. In particular, I will showcase an application of our approach to scalar field theory for calculating entanglement entropy, as well as an application to SU(3) gauge theory aimed at addressing topological freezing.

Speaker: Elia Cellini (University of Turin / INFN Turin)

talkETH_EC.pdf

14:25 → 15:00 Application of generative models on SU(3) gauge theories in 4 dimension

Speaker: Javad Komijani (ETH Zurich)

ml4lattice.pdf

15:00 → 15:40 Coffee break

15:40 → 16:15 Score-Based Diffusion Models for Lattice Gauge Theory

Speaker: Octavio Vega

ML4LQCD_ETH___Talk-2.pdf

16:15 → 17:15 Discussion

Speakers: Marina Krstic Marinkovic (ETH Zurich), Prof. Sinead Ryan (Trinity College Dublin)

ML4Lattice_Wednesday_Discussion.key

19:00 → 22:00 Workshop Dinner

Thursday 12 June

10:00 → 10:50 Progress in learning contour deformations for observables

Speaker: William Detmold

2025_Zurich_ML.pdf

10:50 → 11:30 Coffee break

11:30 → 12:20 Diffusion models for lattice field theory

Speaker: Prof. Gert Aarts (Swansea University)

talkETH-june2025.pdf

12:20 → 13:50 Lunch

13:50 → 14:40 Machine Learning Low-D Systems

I discuss how machine learning is used in stochastic simulations of low-D strongly correlated systems. In particular, I show how machine learning is used to alleviate the numerical sign problem in systems that are doped and/or non-bipartite. I further discuss how flow-based generative models can be used to address ergodicity issues in low-D simulations. Finally, I argue that low-D systems offer a great testbed for testing novel algorithms that could potentially be used in lattice gauge theory simulations.

Speaker: Prof. Thomas Luu (Forshungszentrum Jülich)

ETHZurich_Luu.key

14:40 → 15:15 Simulations of machine-learned RG-improved SU(3) gauge action

Speaker: Kieran Holland

FPA_LGE-CNN_holland_zurich.pdf

15:15 → 15:45 Coffee break

15:45 → 16:20 Efficient Multilevel Sampling of Lattice Field Theory Near Criticality

We present a hierarchical generative framework for efficient sampling of scalar field configurations near criticality. The method leverages a multiscale structure where coarse and intermediate fields are sampled via conditionally constructed Gaussian Mixture Models (GMMs). Normalizing Flows (NFs) refine these samples through invertible transformations that match the target distribution. This approach enables high Effective Sample Size (ESS) and mitigates critical slowing down. Our results demonstrate improved scalability and accuracy over traditional and superresolution-based methods.

Speaker: Ankur Singha

Multilevel_Ankur.pdf

Multilevel_Ankur.pptx

Multilevel_Ankur_slides

Multilevel_Ankur_slides

16:20 → 17:10 Stochastic automatic differentiation for Lattice QCD

Speaker: Alberto Ramos Martinez (Univ. of Valencia and CSIC (ES))

ZurichML.pdf

17:10 → 17:55 Discussion

Speaker: Andreas Kronfeld (Fermi National Accelerator Lab. (US))

ml4lttice-PD3.pdf

Friday 13 June

10:00 → 10:35 Random Matrix Theory for a nano-GPT

Speaker: Ouraman Hajizadeh (Independent Researcher)

rmt_nano_GPT.pdf

10:35 → 11:25 Optimized Multi-Grid Solvers: Power, Precision, and (Matrix) Products

Speaker: Evan Weinberg (NVIDIA Corporation)

optimized_mg_solvers_pdf_v3.pdf

11:25 → 11:55 Coffee break

11:55 → 12:45 Recent Performance Using MILC and QUDA on Various GPU Systems

I plan to show performance data for the MILC code calling QUDA
for configuration generation. There will be a combination of
benchmark runs and production runs on various systems. I will
also show some recent results for deflation, multigrid, and
gauge flow performed by Leon Hostetler with the assitance of Evan
Weinberg. If there is time I will discuss some work to prepare
for a new ensemble with a lattice spacing of about 0.03 fm and
physically light quarks.

Speaker: Steven Gottlieb

Gottlieb_MachineLearing_ETH2025.pdf

12:45 → 13:00 Closing

Speaker: Marina Krstic Marinkovic (ETH Zurich)

13:00 → 14:00 Lunch