AI Winter School 2025 - CFPU/Brown University Department of Physics

Jan 13, 2025, 1:00 PM → Jan 16, 2025, 4:30 PM America/New_York

Richard Gaitskell

Welcome to the 2025 AI Winter School, hosted by the Center for the Fundamental Physics of the Universe, Brown University/Department of Physics

 

Recordings of each module are accessible here on the Brown Physics YouTube channel. (Last year's modules can be viewed here.)

 

This 4-day Winter School is open to interested graduate students, post-doctoral researchers and advanced undergraduates from any institution. Each module and following workshop are intended to stand alone; participants may sign up for as many modules as they choose. 

All modules will take place virtually. There is no registration fee.

AI Winter School Goals

The goal of this Winter School is to offer examples of the direct applications of machine learning methods to a range of problems inspired by the use of artificial intelligence in physics. We will provide hands-on experiences with machine learning tools to help reduce the barriers to engagement with these essential tools for physics research. 

Hosted by the Brown Center for the Fundamental Physics of the Universe, this year's program will consist of six modules comprised of a lecture session delivered by Brown Physics faculty members and industry experts, followed by practice workshop sessions for participants with direction in the use of machine learning tools.

To learn more about the graduate programs in Physics at Brown, please visit the Physics Department website. 

This year's program includes:

Roundtable discussion: "What do we need AI to be capable of to really propel physics forward?"

Featuring: Jim Halverson (Northeastern University), Max Tegmark (MIT), and Kyle Cranmer (University of Wisconsin-Madison); moderated by Richard Gaitskell (Brown University).

• Module: Introductory Module
• Presenter: Shawn Dubey (Brown University)

 

• Module: Development and Deployment of Graph Neural Networks in Particle Physics
• Presenter: Loukas Gouskos and Lazar Novakovic (Brown University)

 

• Module: Physics-Inspired Operator Learning for Inverse Scattering with Application to Ground Penetrating Radar
• Presenter: Yanting Ma (Mitsubishi Electric Research Laboratories)

 

• Module: Unsupervised Learning to Find Interacting and Starburst Galaxies
• Presenter: Ian Dell'Antonio and Philip LaDuca (Brown University)

 

• Module: Overview of Large Language Models (LLMs) and RAG
• Presenter: James Verbus (Senior Staff Software Engineer, Machine Learning, LinkedIn)

 

• Module:  Auto-Encoders for Data Compression in Dark Matter Direct Detection Experiments
• Presenters: Shawn Dubey and Woody Hulse (Brown University)

 

• Module: Generative AI, Agents, and Industry Applications 
• Presenter: Alexis Johnson (Deloitte SFL Scientific), Michael Luk (Deloitte, SFL Scientific)

 

Link to short descriptions of all modules

Register for access to all materials

Registration with a Gmail or Google-backed email account is required for full access to all AI Winter School materials including notebooks. Please note that we will use your registration email address to grant access to all materials. Full access to all materials will be available starting January 12. The AI School will make use of Google Colab for all the workbook examples (https://colab.research.google.com/).

If you do not have a Gmail or Google-backed account, you will need to register one.  You can create an account just for this event or perhaps follow the directions at the following link to use an existing email.

Using an Existing Email with Google

NOTE: Module hands-on material will be on a Google shared drive.  Access to this shared drive will be granted to those that have registered through this Indico site with their GMail or Google-backed email account.  Due to the high level of interest this year, registrants will be granted this access via a dedicated Google group to which their registered email will be added.  Access to the shared drive will not be given directly to any registrant's individual account.  If you did not register with a GMail or Google-backed account, please modify or re-register through this site with a GMail or Google-backed account.  If you just registered, please allow a little time for your registered email to be propagated to the Google group.

The Shared Google Drive will continue to be made available beyond the end of the School so there is no need to copy all the contents. Just your working notebook.

The time zone for the school will only be the US East Coast time zone.

Join the AI School Slack for discussion, trouble-shooting, and questions

An invite link provided here to the to Support Slack. 

Citation Guide:

If you chose to use any of the material from this workshop in your own examples, or workbooks, please make a reference back to our workshop, url, and organizers: 
 

AI Winter School - The Center for the Fundamental Physics of the Universe / Department of Physics, Brown University
https://indico.physics.brown.edu/e/AIWinterSchool2025
 

Organizing faculty: Richard Gaitskell, Director; Ian Dell'Antonio, Associate Director

Monday, January 13

1:00 PM → 2:00 PM Roundtable: What do we need AI to be capable of to really propel physics forward?

Join via Zoom here

Conveners: Jim Halverson (Northeastern University), Kyle Cranmer (University of Wisconsin-Madison), Max Tegmark (MIT), Richard Gaitskell (Brown University)

2:30 PM → 4:00 PM (Module 1) Introductory module

Join via Zoom here

This module is the introductory module to the virtual AI Winter School. It will consist of a short introductory lecture to AI/ML and then a brief hands-on session that will introduce participants to the basics of using Google Colab, the required platform for this winter school. This module is particularly geared toward those with no knowledge of AI/ML and is recommended as a prerequisite to other modules for those coming in with no background in machine learning.

Convener: Shawn Dubey

Tuesday, January 14

10:00 AM → 12:30 PM (Module 2) Graph Neural Networks in Particle Physics

Join via Zoom here

In recent years, Graph Neural Networks (GNNs) have emerged as a transformative tool in particle physics, offering a powerful framework for analyzing complex, non-Euclidean data structures such as particle interactions and detector outputs. This session will provide an exploration of GNNs, bridging their theoretical foundations with practical applications in particle physics. We will highlight their relevance to cutting-edge research at CERN (e.g., ATLAS and CMS), where GNNs are revolutionizing tasks such as particle identification, event classification, and detector signal reconstruction. The hands-on section will feature two critical tasks that showcase the importance of GNNs: a supervised task (particle identification) and a weakly supervised task (particle shower reconstruction). We will use Python-based frameworks and tools (i.e., PyTorch) widely used for GNN development and deployment.

Conveners: Lazar Novakovic, Loukas Gouskos (Brown University)

2:00 PM → 4:30 PM (Module 3) Physics-Inspired Operator Learning for Inverse Scattering with Application to Ground Penetrating Radar

Join via Zoom here

Convener: Yanting Ma (Mitsubishi Electric Research Laboratories (MERL))

Wednesday, January 15

10:00 AM → 12:30 PM (Module 4) Unsupervised Learning to Find Interacting and Starburst Galaxies

Join via Zoom here

Galaxy interactions are laboratories for dark matter physics, star formation and galaxy evolution. Interacting and starburst galaxies can be detected in deep imaging surveys but represent a small fraction of the tens of millions of galaxies. Furthermore, interactions occur in many different scenarios and have different signatures, such as tidal tails, resonances or rings, and disrupted disks, which complicate the training of supervised learning networks. A promising technique for selecting candidate interacting galaxies involves constructing feature-distance maps to organize the images of galaxies, then finding groupings in feature space. In this module, this unsupervised technique will be demonstrated both standalone and in conjunction with a supervised network as a pre-filter for potentially interesting features, based on galaxy images taken from the Dark Energy Camera in Chile. Participants will come away with an understanding of the broad applicability of the technique beyond optical astronomy, to general anomaly detection in image collections.

Convener: Ian Dell'Antonio (Brown University)

2:00 PM → 4:30 PM (Module 5) Overview of Large Language Models (LLMs) and RAG

Join via Zoom here

This module will provide an exploration of LLM tools and techniques, including using OpenAI API, an open LLaMA model, retrieval-augmented generation (RAG) for improving AI system performance with external data, and strategies for safeguarding LLM implementations with LLaMa guard. We will set up an LLM using both OpenAI’s API and a locally installed LLaMa model, build a basic RAG system to demonstrate how external data can be leveraged to enhance AI outputs, and show how LLaMa guard can moderate input prompts to ensure safe and appropriate responses in practical applications. Participants will leave with a familiarity of and the ability to use these LLM tech stack components.

Convener: James Verbus

Thursday, January 16

10:00 AM → 12:30 PM (Module 6) Auto-Encoders for Data Compression in Dark Matter Direct Detection Experiments

Join via Zoom here

The LUX-ZEPLIN (LZ) dark matter direct detection experiment searches directly for dark matter particle interactions in a 10-tonne liquid xenon target. LZ collects about 50 MBytes per second of detector data across its 494 photomultiplier tubes (PMTs), or around 1.5 PBytes (1.5 million GB) per year. The utility of this data naturally becomes constrained by its scale–to efficiently analyze this quantity of information, we need to achieve compressed, rich representations of the data. Artificial neural networks (ANNs) have long been applied as an effective way to classify, model, and reason over large quantities of data due to their ability to learn over vast hyperspaces. Autoencoders (AEs), deep neural networks composed of encoder and decoder components, are one method used in particular for the task of compressing data. In this module you will learn about the formulation, benefits, and limitations of AEs and leverage them for applications to LZ data compression, which also can be useful in other areas of research.

Conveners: Shawn Dubey, Woody Hulse

2:00 PM → 4:30 PM (Module 7) Generative AI with industry applications

Join via Zoom here

Convener: Michael Luk