Machine Learning approaches in Lattice QCD - An interdisciplinary exchange
Institute for Advanced Study of the Technische Universität München
A workshop for an interdisciplinary exchange for accelerating developments of machine learning techniques in Lattice QCD with experts in Machine learning, Lattice QCD, and other related fields.
This is a Covid safe workshop: FFP2 masks will be required.
Please remember to bring with you FFP2 masks
The conference is broadcasted in zoom if you want to join, contact the organizers via email.
-
-
8:30 AM
→
9:10 AM
Registration 40m
-
9:10 AM
→
9:25 AM
Welcome and aims of the workshop 15mSpeaker: Nora Brambilla
- 9:25 AM → 9:55 AM
-
9:55 AM
→
10:45 AM
Generative flow methods
Talks and panel discussion
Convener: Sinead Ryan-
9:55 AM
Aspects of scaling and scalability for flow-based samplers 25mSpeaker: Daniel Hackett
- 10:20 AM
-
9:55 AM
-
10:45 AM
→
11:15 AM
Coffee break 30m
-
11:15 AM
→
12:40 PM
Generative flow methods
Talks and panel discussion
Convener: Sinead Ryan- 11:15 AM
-
11:40 AM
Stochastic normalizing flows for lattice field theory 20mSpeaker: Elia Cellini
-
12:00 PM
Panel discussion 40mSpeakers: Chair: Sinead Ryan, Members: Simone Bacchio, Daniel Hackett, Maria Paola Lombardo, Alberto Ramos
-
1:00 PM
→
2:30 PM
Lunch 1h 30m
-
2:30 PM
→
4:00 PM
ML for particle physics
Talks and panel discussion
Convener: Lukas Heinrich-
2:30 PM
Discovering mathematical structures with neural networks 20mSpeaker: Sven Krippendorf
- 2:50 PM
-
3:10 PM
Improving Normalizing Flows to Sample from Boltzmann Distributions 20mSpeaker: Vincent Stimper
- 3:30 PM
-
2:30 PM
-
4:00 PM
→
4:30 PM
Coffee break 30m
-
4:30 PM
→
6:00 PM
ML for phase transitions and sign problem mitigation: Contour deformation methods for noise reduction
Talks and panel discussion
Convener: Will Detmold-
4:30 PM
Signal-to-noise improvement with contour deformations 20mSpeaker: Gurtej Kanwar
-
4:50 PM
Complex normalizing flows and subtractions for sign problems 20m Zoom
Zoom
https://tum-conf.zoom.us/j/68490931000 Meeting-ID: 684 9093 1000 Passcode: ML4LatticeSpeaker: Yukari Yamauchi -
5:10 PM
Applying Complex Valued Neural Networks to the Hubbard Model Sign Problem: A Survey and Case Study 20mSpeaker: Marcel Rodekamp
-
5:30 PM
Panel discussion: 1. maximum improvement in <sign> 2. ML towards such a maximal improvement 3. Contour deformation for specific observables 4. Translation between observables/theories 30m
1) Is there a way to determine the maximum improvement in <sign> or in S-to-N that is achievable through contour deformations and related
approaches?2) How can ML be targeted towards such a maximal improvement?
3) Are there features of specific observables that make contour
deformation particularly effective or ineffective?4) How do contour deformations for one observable/theory translate to other observables/theories?
Speakers: Chair: Will Detmold, Members: Gurtej Kanwar, Andrei Alexandru (The George Washington University), Thomas Luu, Yukari Yamauchi
-
4:30 PM
-
8:30 AM
→
9:10 AM
-
-
9:50 AM
→
10:20 AM
Model-Independent Learning of Quantum Phases of Matter with Quantum Convolutional Neural Networks 30mSpeaker: Frank Pollmann
- 10:20 AM → 10:45 AM
-
10:45 AM
→
11:15 AM
Coffee break 30m
-
11:15 AM
→
1:00 PM
Gauge field generation
Talks and panel discussion
Convener: Phiala Shanahan-
11:15 AM
Simulation of the 2D Schwinger Model via machine-learned flows in Global Correction steps 25mSpeaker: Jacob Finkenrath
-
11:40 AM
Variational Autoregressive Networks for Information Theory 20m Zoom
Zoom
https://tum-conf.zoom.us/j/68490931000 Meeting-ID: 684 9093 1000 Passcode: ML4LatticeSpeaker: Tomasz Stebel -
12:00 PM
Machine-learning-assisted Monte Carlo fails at sampling computationally hard problems 20mSpeaker: Jeanne Trinquier
-
12:20 PM
Panel discussion: 1. Utility of ML for lattice 2. application-specific features of gauge field generation 3. Biggest challenges 4. Applications at an interesting scale 40m
• Temperature check: for each panelist, in 2-5 words, how optimistic or pessimistic do you feel about the utility of ML in accelerating lattice field generation in your context.
• What are the most application-specific features of the LQCD gauge field generation problem i.e., in what important ways is it different from applications in other areas of physics, and how does that impact the way field generation algorithms must be designed?
• What are the biggest challenges to lattice field generation accelerated by ML?
• Circling back to the start: what are your reasons for optimism or pessimism in reaching applications at an interesting scale?Speakers: Chair: Phiala Shanahan, Members: Fernando Romero Lopez, Jacob Finkenrath, Jeanne Trinquier
-
11:15 AM
-
1:00 PM
→
1:15 PM
Group Photo 15m
-
1:15 PM
→
2:30 PM
Lunch 1h 15m
-
2:30 PM
→
4:00 PM
ML for phase transitions and sign problem mitigation
Talks and panel discussion
Convener: Maria-Paola Lombardo-
2:30 PM
Machine learning for quantum field theories with a sign problem 20mSpeaker: Andrei Alexandru
-
2:50 PM
A machine learning approach to the classification of phase transitions in many flavor QCD 20mSpeaker: Marius Neumann
-
3:10 PM
Teaching how to teach: defining learning samples to detect phase transitions 20mSpeaker: Francesco Di Renzo
-
3:30 PM
Panel discussion: 1. Competitivity of ML for Phase identification 2. ML for Hypothesised Phases 3. Potential of ML for sign porblem 30m
Panel discussion addressing the three main questions:
--Where do we stand with the identification of phases with ML?
Do you feel ML could become competitive with standard techniques for the quantitative studies of known phase transitions i.e. order of the transition at large Nf, chiral limit, etc.--Which are your views on the need/concrete possibility of identifying hypothesised, not yet detected phase transitions - in particular confinement, axial restoration, anomalous threshold in the QGP
--Can you summarise the status of the sign problem (without ML) (finite baryon density/real time) and how do you assess the potentiality of ML to mitigate it?
Speaker: Chair: Maria Paola Lombardo
-
2:30 PM
-
4:00 PM
→
4:30 PM
Coffee break 30m
-
4:30 PM
→
6:00 PM
Spectral ReconstructionConvener: Gert Aarts
-
4:30 PM
Brief introduction on spectral functions and their computation 10mSpeaker: Gert Aarts
- 4:40 PM
-
5:10 PM
Spectral reconstruction with neural networks 30mSpeakers: Kai Zhou, Lingxiao Wang
-
5:40 PM
Learning regulators for spectral reconstruction 10mSpeaker: Alexander Rothkopf
-
5:50 PM
Panel discussion 10mSpeaker: Chair: Gert Aarts
-
4:30 PM
-
9:50 AM
→
10:20 AM
-
-
8:45 AM
→
9:00 AM
Introducing Munich Data Science Institute 15mSpeaker: Sylvia Kortüm
-
9:00 AM
→
10:40 AM
EFT, Lattice and ML
Talks and panel discussion
Convener: Nora Brambilla (Physik Department, TU Munich)- 9:00 AM
-
9:40 AM
An analysis of Bayesian estimates for missing higher orders in perturbative calculations 30mSpeaker: Aleksas Mazeliauskas
- 10:10 AM
-
10:40 AM
→
11:10 AM
Coffee break 30m
-
11:10 AM
→
1:00 PM
EFT, Lattice and ML
Talks and panel discussion
Convener: Nora Brambilla (Physik Department, TU Munich)-
11:10 AM
Finite-volume pionless EFT for multi-nucleon systems with differential programming 30mSpeaker: Fernando Romero-Lopez
-
11:40 AM
Panel discussion: 1. What can EFTs do for ML4LATTICE? 2. What ML4LATTICE could do for EFTs? 1h 10m
--what can EFTs do for ML4LATTICE?
--what ML4LATTICE could do for EFTs?Speakers: Chair: Nora Brambilla, Members: Marina Marinkovic, Phiala Shanahan, William Detmold, Aleksas Mazeliauskas
-
11:10 AM
-
1:00 PM
→
2:30 PM
Lunch 1h 30m
-
2:30 PM
→
4:35 PM
ML as component of exact algorithms
Talks and panel discussion
Convener: Phiala Shanahan-
2:30 PM
Building Transport Maps and Making Good Use of Them 30mSpeaker: Michael Albergo
- 3:00 PM
-
3:25 PM
Symbolic Distillation of Neural Networks 30m Zoom
Zoom
https://tum-conf.zoom.us/j/68490931000 Meeting-ID: 684 9093 1000 Passcode: ML4LatticeSpeaker: Miles Cranmer -
3:55 PM
Panel discussion: 1. Meaning of "exact" algorithms 2. in-principle vs in-practice exactness 3. applications for exactness 40m
• What does it mean to have an ML-accelerated algorithm that is “exact”?
Discuss the distinction between exact algs, interpretable algs, and algs that allow error propagation.
• What are the differences between in-principle and in-practice exactness? Are they important?
• In what applications (both in LQFT and drawing parallels to other ares in physics) is it important to guarantee exactness in ML-accelerated algorithms, and where is it unnecessary, impossible, or worth sacrificing?Speakers: Chair: Phiala Shanahan, Members: Michael Albergo, Andrei Alexandru, Miles Cranmer
-
2:30 PM
-
4:35 PM
→
6:00 PM
Reception 1h 25m
-
6:00 PM
→
7:00 PM
How to accelerate gauge field field generation using flow-based and hybrid models 1hSpeaker: Phiala Shanahan
-
8:45 AM
→
9:00 AM
-
-
9:30 AM
→
10:00 AM
ML for particle physics 30mSpeaker: Lukas Heinrich
-
10:00 AM
→
10:30 AM
Complex Langevin real-time simulations and ML 30mSpeaker: Alexander Rothkopf
-
10:30 AM
→
10:50 AM
ML for physical interpretation of lattice results
Talks and panel discussion
Convener: Alexander Rothkopf- 10:30 AM
-
10:50 AM
→
11:15 AM
Coffee Break 25m
-
11:15 AM
→
1:00 PM
ML for physical interpretation of lattice results
Talks and panel discussion
Convener: Alexander Rothkopf-
11:15 AM
Physical Concepts from Neural Networks with Two Inputs 20mSpeaker: Sebastian Wetzel
- 11:35 AM
-
12:00 PM
Panel discussion on ML for physics interpretation 1h
- Alexander Rothkopf: Setting the stage – ML for physics interpretation
- Sebastian Wetzel: ML for identifying order parameters & effective d.o.f. in the context of phase transitions
- Matteo Favoni: ML for identifying defects and effective d.o.f. in the context of confinement
- Fernando Romero Lopez: ML for extraction of emergent d.o.f. in the context of scattering resonance
- Miles Cranmer: ML for inference of governing equations in the context of orbital mechanics (remote)
- Discussion with the audience
Speakers: Chair: Alexander Rothkopf, Members: Sebastian Wetzel, Matteo Favoni, Fernando Romero Lopez, Miles Cranmer
-
11:15 AM
-
1:00 PM
→
2:30 PM
Lunch 1h 30m
-
2:30 PM
→
3:40 PM
Symmetry equivariant neural networks/informed neural networks
Talks and panel discussion
Convener: Phiala Shanahan-
2:30 PM
Generation of gauge field configurations with equivariant neural networks 20mSpeaker: Matteo Favoni
- 2:50 PM
-
3:05 PM
Panel discussion: 1. Symmetry-equivariance vs computational brute force 2. Computational challenges of gauge equivariance 35m
• There is ongoing debate in the ML community about the value of trying to incorporate domain knowledge into architectures vs using computational brute force. What guides your intuition for the value of building symmetry-equivariant architectures? In your experience when is symmetry-equivariance valuable and when is it not?
• What makes gauge equivariance in particular challenging from a computational efficiency perspective?Speakers: Chair: Phiala Shanahan, Members: Gurtej Kanwar, Christoph Lehner, Matteo Favoni
-
2:30 PM
-
3:40 PM
→
6:40 PM
Poster session (With food and drinks)
-
9:30 AM
→
10:00 AM
-
- 9:00 AM → 9:30 AM
-
9:30 AM
→
11:00 AM
Inverse Problem
Talks and panel discussion
-
9:30 AM
Machine learning hadron spectral functions in Lattice QCD 30mSpeaker: Gabor Papp
-
10:00 AM
R-ratio from Lattice QCD 30mSpeaker: Jian Liang
-
10:30 AM
Inverse problem solving in nuclear physics with deep learning 30mSpeaker: Kai Zhou
-
9:30 AM
-
11:00 AM
→
11:30 AM
Coffee break 30m
-
11:30 AM
→
12:30 PM
Inverse Problem
Talks and panel discussion
Convener: Andreas Kronfeld-
11:30 AM
On the extraction of hadronic spectral densities from Euclidean correlators 30mSpeaker: Alessandro De Santis
- 12:00 PM
-
11:30 AM
-
12:30 PM
→
2:00 PM
Lunch 1h 30m
-
2:00 PM
→
3:15 PM
EFT, Lattice and ML
Talks and panel discussion
-
2:00 PM
Machine learning for symbolic computation in high energy physics 25m zoom
zoom
https://tum-conf.zoom.us/j/68490931000 Meeting-ID: 684 9093 1000 Passcode: ML4LatticeSpeaker: Abdulhakim Alnuqayqdan -
2:25 PM
Alleviating the Sign Problem via Contour Deformation and Machine Learning 25mSpeaker: Thomas Luu
-
2:50 PM
Gauge-equivariant neural networks as preconditioners in lattice QCD 25mSpeaker: Christoph Lehner
-
2:00 PM
-
3:15 PM
→
3:40 PM
Coffee break 25m
-
3:40 PM
→
5:00 PM
EFT, Lattice and ML
Talks and panel discussion
-
3:40 PM
Deep Learning and the Standard Model: a philosophy of science perspective 30mSpeaker: Luigi Scorzato
-
4:10 PM
The route towards phase transition recognition in Lattice Gauge Theories 30mSpeaker: Andreas Athenodorou
-
4:40 PM
Search for Exotic Higgs Bosons using Quantum Machine Learning 20mSpeaker: Nakul Aggarwal
-
3:40 PM
-
5:00 PM
→
5:15 PM
Summary 15mSpeaker: Nora Brambilla