May 23, 2024

May 23, 2024

Special Seminar: Thursday, May 23, 1.05 pm Eastern Special Time

• Speaker: Benjamin Nachman LBL DOE Laboratory
• Title: Generative AI for the Pursuit of Fundamental Interactions
• Abstract: Deep generative models are being developed, adapted, and deployed to a number of challenges in particle physics. In this talk, I will survey some of these topics, including surrogate modeling for slow physics-based simulations, anomaly detection to search for new particles, and simulation-based inference.
• Presentation: GenerativeAISpring2024.pdf (open)
• Recording of Session: VisibleScience Working Group (2024-05-23 10_03 GMT-7).mp4 (skip first 12 minutes 30 seconds)

Quoted Papers

• A Living Review of Machine Learning for Particle Physics: https://iml-wg.github.io/HEPML-LivingReview/
• Calorimeter Simulation Accelerating Science with Generative Adversarial Networks: An Application to 3D Particle Showers in Multi-Layer Calorimeters https://arxiv.org/abs/1705.02355
• Kaggle Fast Calorimeter Simulation Challenge 2022 https://calochallenge.github.io/homepage/
• CaloScore: (Diffusion model for Calorimeter Simulation from Nachman) Comparison of Point Cloud and Image-based Models for Calorimeter Fast Simulation https://arxiv.org/abs/2307.04780
• CaloScore v2: Single-shot Calorimeter Shower Simulation with Diffusion Models https://arxiv.org/abs/2308.03847
• Analyze LIGO Gravitational wave data:Real-time gravitational-wave science with neural posterior estimation https://arxiv.org/abs/2106.12594
• Hadronization Towards a Deep Learning Model for Hadronization https://arxiv.org/abs/2203.12660 https://journals.aps.org/prd/abstract/10.1103/PhysRevD.106.096020
• Hadronization Fitting a Deep Generative Hadronization Model https://arxiv.org/abs/2305.17169
• Hadronization Integrating Particle Flavor into Deep Learning Models for Hadronization https://arxiv.org/abs/2312.08453
• Deconvolution compare NF and diffusion model Improving Generative Model-based Unfolding with Schrödinger Bridges https://arxiv.org/abs/2308.12351 https://journals.aps.org/prd/abstract/10.1103/PhysRevD.109.076011
• Deconvolution The Landscape of Unfolding with Machine Learning https://arxiv.org/abs/2404.18807
• Anomaly Detection Classifying Anomalies THrough Outer Density Estimation (CATHODE) https://arxiv.org/abs/2109.00546
• Foundation Model OmniLearn: A Method to Simultaneously Facilitate All Jet Physics Tasks https://arxiv.org/abs/2404.16091

Notes on Presentation

• 21 people attended this excellent talk
• The presentation started by describing fundamental physics and some places where generative AI is used
• Initially Caloscore and the Kaggle challenge were described
• The ata sparse and so use point clouds
• The Benjamin described a bolder goal to describe the full process of simulation, hadronization and reconstruction. Here one has a point cloud in both input and output
• Best to map quarks and gluons to observed particles as hadronization uncertain.
• FHadronization and off line problems speed is not important whereas critical for CaloScore type problem.
• This leads to Unfolding with Schrödinger Bridge and other methods. Both GANs and diffusion models are used. Normalizing Flows similar
• Described a NIPS Data challenge for anomaly detection. FAIR Universe HiggsML Challenge
• Slide 19 show how this can lead to improved identification of new signals which are reliable when classical methods would give a lower significance than the chassical “five sigma” usually looked for Why do physicists mention “five sigma” in their results? | CERN.
• Slide 20 describes possible use of Differentiable simulations in hadronization problem.
• Benjamin described recent Foundation model work
• Gans are hardest to train but flexible so use in hadronization; also GANs faster than diffusion models
• Normalizing Flows give you density whereas diffusion model gives score only
• Tom Gibbs noted the related NVIDIA GTC presentation GTC_2024_HEP by Vinicius M. Mikuni, also from LBL.