1 of 70

AI in HEP: A3D3 and IAIFI

Shih-Chieh Hsu

University of Washington

USLUA Dec 18 2025�

Slides featuring Phil Harris input

1

PHY-2117997

https://a3d3.ai/

DE-SC-0023527

2 of 70

2

?

Deep Learning �revolution

2020

2022 2024

GenAI

revolution

ChatGPT

Agentic AI

AlphaFold

3 of 70

3

700PB/year

The Square Kilometre Array (SKA)

HL-LHC

100PB/year~EB/year

Next-Generation Big Data Frontiers

700PB/year

200PB/year

Advanced Photon Source (APS)

Linac Coherent Light Source II �(LCLS-II)

100PB~EB/year

Exabyte-scale Big Data Challenge

4 of 70

AI/ML for HEP�Challenge and Opportunities

Growing data complexity and advanced detectors
Need for real-time workflows
AI for real-time event selection

Standardize detector data for AI
DAQ/calibration integration
Open science and interoperability

5 of 70

NSF Institutes for HEP

https://iris-hep.org/

Three major NSF-funded institutes, each addressing unique aspects of computational and AI-driven research at the intersection of physics and data science.

aims to meet the software and computing challenges posed by the High Luminosity LHC (HL-LHC), developing state-of-the-art cyberinfrastructure and acting as a community-wide hub for software R&D in high energy physics.

https://a3d3.ai/

focused on fusing foundational physics principles with cutting-edge AI approaches to tackle challenging problems in physics and galvanize innovation in trustworthy AI.

https://iaifi.org/

targets real-time AI solutions for large, complex datasets across high energy physics, multi-messenger astrophysics, and systems neuroscience, integrating customized AI with advanced hardware acceleration.

6 of 70

NSF HDR Institute: Accelerated Artificial Intelligence Algorithms for Data-Driven Discovery

6

Mission: To advance scientific discovery through real-time AI applications at scale.��
Vision: To empower researchers with the knowledge and tools for effective real-time AI use across scientific fields.�
Starting 2021: 15M for 5 years

7 of 70

A3D3 Members

7

Founding Members: 128 members from 10 institutions, including 21 senior personnel, 16 postdoc, 82 students.

(4 Early career awardees, 2 Sloan fellows, 1 AAAS)

Affiliate Members: 40 member from 10 institutions, consisting of 20 faculty/staff, 5 postdoc, 15 students.

(affiliate faculty including 4 A3D3 alumni )

Global partners

NSF AREAS Award

8 of 70

A3D3 Cross-discipline

8

HEP

(13)

MMA

(7)

Neuros

(5)

CS/EE�(9)

Hsu

PI/Director

Harris

co-PI

Neubauer

co-PI

Liu

Duarte

Hauck

Li

Han

Chen

Riedel

Orsborn

Shlizerman

Dadarlat Makin

Coughlin

co-PI

Scholberg

co-PI

Graham

Katsavounidis

20 out of�34 senior personnel are HEP+MMA

Ju

Lai

Rankin*

Sravan*

Li

Aarastad

Sun

Gonski

Carlsen

Cremonesi

DiPetrillo

Cavanaugh

Yu

Buat

Liu*

Li

Khoda*�

9 of 70

High Energy Physics

Enabling the potential discovery of new elementary particle physics by developing AI algorithms to process data with sub-microsecond ultra-low latency and 1 Petabit/sec date rate (ATLAS, CMS)

9

40M collisions/seconds

10 of 70

Multi-Messenger Astrophysics

Enable rapid identification and follow-up of gravitational-wave events using LIGO and optical facilities (ZTF/LSST).
Improve real-time detection and localization of supernova neutrinos with IceCube and DUNE.

10

11 of 70

Neurosciences

Improve our understanding of how neural activity gives rise to behavior through real-time interventions and high throughput data exploration

11

Brain

Behavior

12 of 70

Innovation

Transdisciplinary Research pursued by integrating expertise from diverse fields, such as high energy physics, multi-messenger astrophysics, neuroscience, computer science and electrical engineering, to tackle common scientific challenges.

Inference-as-a-Service for science: Client applications use standardized APIs to simplify hardware and software details, enabling seamless integration of inference capabilities across heterogeneous scientific environments.

Gunny et al Nature Astronomy (2022)

Exploring Common Analysis Challenges: Researchers employ deep learning for anomaly detection and forecasting to study complex phenomena across disciplines using shared time series data.

R. Raikman et al MLST 5 025020 (2024)

Li et. al. NeurIPS 2023 [c43]

ML4GW/HERMES

Low-latency AI deployment: �Efficient hardware and infrastructure are developed to deploy and accelerate ML/AI algorithms across various scientific domains.

https://github.com/fastmachinelearning/hls4ml

13 of 70

Cross-disciplinary collaboration by publications

13

HAC

HEP

MMA

Neuroscience

Fostering and strengthening collaboration across all focus areas.

14 of 70

Cochran-Branson

15 of 70

HEP: ANOMALY DETECTION AT 40 MHZ

15

Four jets with 𝑝) > 50 GeV
three electrons and a muon with 𝑝T > 30 GeV, and
𝐸Tmiss of 215 GeV

16 of 70

MMA: Aframe + AMPLFI deployed Summer 2025!

Made the first Compact Binary Coalescence discovery using NNs

S250904cv, S250904br, S250904ae, S250901cb, S250830bp, S250830m
First alert from Aframe.

16

Posterior

Likelihood

Prior

In Likelihood-free Inference:

Learn the distribution from simulations

NN approximator

17 of 70

Education and Outreach

17

J. Duarte at ICML 2024 (by P. Li)

Education and Outreach

21 A3D3 seminars

7 Tutorials

10 Course curriculum

25 Undergrad researchers

5 High schoolers

4 K-12 Science Fest

We design and implement tailored activities for diverse educational programs, ranging from K-12 to Expert levels.

18 of 70

Equity and Career

18

Equity and Career

12 Postbac Fellows

(Increasing applicants

Year 1: 13
Year 2: 55
Year 3: 87)

1 Career development event

14 Mentor-mentee pairs

4 Trainee-led events� (Monthly seminar, All-hands, Town hall)

6 STEM diversity events

Womxn
SACNAS 2024
NSHP/NSBP 2024

DEI-Climate survey

A3D3 is happy to share the booth with HDR Ecosystem.

19 of 70

Community Engagement

19

Community Engagement

29 Affiliate (launched in May 2024)

2 NSF Partnership � 6 Industry/HPC collaboration

6 HDR Ecosystem events

(conf, workshop)

4 International Fast ML events� (conf, workshops)

AREAS: Accelerating Research and Education in AI for Science

To grow and sustain the hls4ml open-source software ecosystem

To advance STEM for minority serving institute

Northwestern, TI-2303700

UIC, NSF PREP 24-514

FastML@ICCAD 2023

HDR Eco. Conf. 2025

20 of 70

Taiwan Tech Connect

June 2025 80 ppl

US Taiwan Tech Connect

250 ppl

Industry Connection

21 of 70

Mike Williams | Phiala Shanahan | Marisa LaFleur | Thomas Bradford

IAIFI Interim Director | IAIFI Interim Deputy Director | IAIFI Managing Director | IAIFI Project Coordinator

NSF Institute for Artificial Intelligence

and Fundamental Interactions (IAIFI)

Slides By: Philip Harris

22 of 70

Deep Learning (AI) + Deep Thinking (Physics) = Deeper Understanding

22

= AI + Physics

Pioneering �interdisciplinary

RESEARCH

Building a dynamic

COMMUNITY

Empowering the �next generation of

TALENT

Tackling two of the greatest mysteries of science through curiosity-driven research: �how our universe works and how intelligence works

22

The philosophical foundation of IAIFI is elegant: combine deep learning—today's most powerful AI paradigm—with deep theoretical and experimental physics to achieve insights neither field could reach alone.

IAIFI operates on three pillars. First is interdisciplinary research: genuine fusion of AI and physics methodologies, not parallel silos. Second is talent development: training the next generation of researchers who are genuinely fluent in both languages—physicists who understand neural network architecture, AI researchers who grasp quantum field theory. Third is community building: creating venues where physicists and AI researchers interact, share problems, and cross-pollinate ideas. This last point is crucial because the vernacular is genuinely different. That cross-language capability is what makes IAIFI's contributions stick in the HEP community.

23 of 70

23

Faculty & Senior Investigators (5/22 are HEP/LIGO)

Affiliates (7/44)

Mike Williams

MIT/ IAIFI Interim Director

HEP+LIGO Involvement in IAIFI

Project Management

Marisa LaFleur

Managing Director

Thomas Bradford

Project Coordinator

IAIFI Fellows (2/7 in experiment)

Karen Dow�Nico Lang

Lauren Saragosa

Yesenia Ortiz

LNS Admin Support

Faculty Senior Investigators: 5; Affiliates: 7; IAIFI Fellows: 2

Phil Harris

MIT

Taritree Wongjirad

Tufts

Carlos Arguëlles-Delgado

Harvard

Lisa Barsotti

MIT

Eluned Smith

MIT

Aram Apyan

Brandeis

Roger Rusack

U. of Minnesota/ IAIFI Visitor

Sam Bright-Thonney

AI for Particle Physics

Gaia Grosso

AI for Particle Physics

Matt LeBlanc

Brown

Pierre-Hugues Beauchemin

Tufts

Sudhir Malik

UPRM

Erik Katsavounidis

MIT

Larry McMahon

Joe Cucinotta

Iling Hong

Ellen Vervaeke

23

Let me establish IAIFI's credentials in our field. Of 22 senior faculty investigators, 5 lead the HEP and gravitational wave programs. You see representatives from MIT, Harvard, Tufts, Brown, UPRM—leading institutions. On the IAIFI Fellow track focused on experimental physics, 2 out of 7 come from our communities.

This isn't ceremonial representation. Phil Harris at MIT, Sam Bright-Thonney in AI for particle physics, Larry McMahon at Harvard, Taritree Wongjirad at Tufts—these are active researchers driving IAIFI's HEP agenda. When you have people of this caliber embedded in both the physics experiments and the AI research institute, you get genuine two-way flow: physicists bring real constraints and data; AI researchers bring novel algorithmic approaches. That bidirectional transfer is what creates impact.

24 of 70

The goal of IAIFI has been on the development of novel AI methods

We aim to create new algorithms and approaches that are robust and interpretable
The focus is much more on the algorithm design and less on the deployment

IAIFI’s impact is the integration of AI into HEP experiments

24

Team part of the first AI anomaly detection paper in CMS. QUAK algorithm developed in IAIFI

Lipschitz networks allow for specific constraints, like positive trigger turn ons, to be built into the architecture. Now being used in LHCb trigger

Impact in HEP

DeepSets Networks have become the baseline for particle based ML methods

24

Now, the concrete outcomes. IAIFI's philosophy from the start has been to develop robust, interpretable AI methods focusing on algorithm innovation—understanding what works and why—rather than premature deployment.

Three examples show this philosophy in action. DeepSets networks, developed within IAIFI, have become the foundational architecture for particle-based machine learning across HEP. We're not talking about one experiment using this—it's the baseline now.

Lipschitz networks enable us to embed specific physics constraints directly into architecture design. For example, ensuring monotonically positive trigger turn-ons isn't a constraint you enforce post-hoc—it's built into the network topology. LHCb is actively using this in their trigger system.

And just recently, IAIFI was part of the first AI anomaly detection publication from CMS, featuring the QUAK algorithm. These are production-ready approaches now running in real detectors.

25 of 70

25

Current Exciting HEP projects

M. Williams: 2509.06855

G.Grosso: 2511.03095

S.Bright-Thonney: 2510.21935

LLMs to parse analysis notes and produce knowledge graphs

Link analysis notes through common methods

Sparker: A completely new neural network architecture built on sparse Gaussian kernels. This method far outperforms all other methods in AI-based anomaly detection, has better interpretability and works on many datasets : LIGO, HEP, ….

Autoscidact: Combine AI anomaly detection strategys with contrastive learning to embed physics knowledge

An automated discovery pipeline from raw inputs

From raw 4-vectors

25

Let me highlight three cutting-edge projects where IAIFI researchers are pushing boundaries right now.

Mike Williams' LLM project tackles an underappreciated bottleneck in HEP: institutional knowledge fragmentation. His team is using large language models to parse physics analysis notes and construct knowledge graphs that automatically link disparate analyses through common methodologies. Imagine the discovery potential—instantly see how different experimental groups approach similar physics, identify methodological innovations, spot gaps. This is meta-science automation.

Gaia Grosso's Sparker is a completely new neural network architecture built on sparse Gaussian kernels. It significantly outperforms existing anomaly detection approaches while being more interpretable—you can actually understand what the network is looking at. Crucially, it works across domains: LHC, LIGO, and beyond.

Sam Bright-Thonney's Autoscidact integrates anomaly detection with contrastive learning to embed physics knowledge into discovery pipelines. The goal: fully automated physics discovery starting from raw four-vector data.

26 of 70

Exciting work in Neutrino Physics

26

C. Arguelles: 2510.01733

A pre-training strategy to make Icecube events more robust to data/MC variations.

Utilizes Transformer and Self-supervised learning to featurize Icecube data.

T. Wongjirad: 2307.13687

A diffusion model to generate liquid argon neutrino events leads to accurate and fast simulations.

Building towards a fully heterogeneous pipeline.

J. Micallef: indico link

A graph NN for DUNE particle track reconstruction and Id

Network integrates multiple detector elements allows for correct/unbiased linking across large spaces

26

While IAIFI's HEP collider program is thriving, their neutrino physics contributions are equally impressive—and often overlooked.

Taritree Wongjirad at Tufts uses generative diffusion models to simulate liquid argon neutrino interactions with remarkable fidelity and speed. This moves toward fully heterogeneous simulation pipelines—a massive efficiency gain for DUNE and other LArTPC experiments.

James Micallef (presenting at NuFact 2025) developed graph neural networks for DUNE track reconstruction that intelligently integrate multiple detector elements, allowing unbiased particle linking across large spatial scales. This solves a real problem in reconstruction ambiguity.

Carlos Arguelles brought pre-training and self-supervised learning strategies to IceCube data, making event reconstruction robust to data-to-simulation discrepancies. Using Transformers to automatically featurize Cherenkov detector data—this is exactly the kind of innovation that advances our experimental capabilities.

27 of 70

IAIFI Research Impact�

27

Advancing physics knowledge and galvanizing AI research innovation

Dark Matter Searches

LHC

IceCube

(& DUNE)

LIGO

Structure Formation

Multi-Messenger Astrophysics

Representation Learning

Robust/ Interpretable AI

Reinforcement Learning

Foundational AI

Many-Body Physics

QFT & String Theory

Standard Model

Theoretical Physics

Experimental Physics

Astrophysics

34 IAIFI collaborations in progress

161 papers on arXiv; �97 papers published

44 coding packages

8,259 citations

Year 5 in Review:

Totals below are from the past year only

Oral presentation at ICLR 2025

NeurIPS Workshop: Spotlight talk & runner-up best paper

Contributions at NeurIPS 2024 main conference

27

Take a broader view, and IAIFI's influence is remarkable. Their work spans foundational AI (representation learning, reinforcement learning, interpretability), theoretical physics (standard model extensions, QFT, many-body systems), experimental physics (LHC, LIGO, IceCube, DUNE), and astrophysics (dark matter, multi-messenger astronomy).

In just the past year: 34 active collaborations. 161 papers on arXiv; 97 published in peer-reviewed venues. 44 new software packages released. Over 8,200 citations. They had contributions at NeurIPS 2024 main conference, an oral presentation at ICLR 2025, and a spotlight talk at NeurIPS workshop that was runner-up best paper.

For context: IAIFI is five years old. This is a young institute producing at the level we associate with much more established centers. That speaks to both the quality of people involved and the fundamental soundness of the model.

28 of 70

Deliverables to the Experiments

IAIFI is developing new AI architectures

The ideas and packages we construct tend to be standalone
Integration into experiments is done on a user basis
Long term sustainability is not a significant issue

An important part of IAIFI is facilitating discussion about critical issues

We aim to provide a portal to communicate methods and ideas

IAIFI summer workshop and summer school
Vernacular between AI and Physics is not the same
Venues aim to bring together Physics and AI to ensure cross-pollination of ideas

Integration with universities and other communities

We are focused on trying to understand how AI+Physics becomes its own subfield
How do we encourage PhDs are joint AI and Physics? (Coursework/Class material)
What are the right ways to communicate across fields (HEP/Astro/Theory/…)

28

You might ask: how does IAIFI ensure its innovations actually reach experiments and stay relevant?

The answer is pragmatic. IAIFI develops AI architectures and software packages designed to be modular and standalone—not tightly coupled to specific experiment infrastructure. This means experimental collaborations can integrate on their own timeline, in their own way. Integration responsibility rests with the experiments, which actually creates long-term sustainability. Experiments own their dependencies, not the other way around.

But IAIFI also plays a critical convening role. They run summer workshops and schools that facilitate discussion about how AI and physics become a unified discipline. They're standardizing vocabulary—because "loss function" and "physics constraint" mean different things to different communities. And they're partnering with universities to build joint PhD programs in AI and Physics, developing coursework that trains researchers to be genuinely bilingual in both fields.

29 of 70

Sustainability

IAIFI is currently up for renewal (we expect feedback relatively soon)

We continue to operate as a normal institution
With all of the partner institutes we have integrated with larger departments

Integration with computer science departments is always a challenge

Software and toolkit sustainability follows standard practices

Since IAIFI works really at the ML design level
Software is released using Github/HuggingFace with datasets
We provide conventional validation tools following computer science practices

Beyond industry tools, we have limited reliance on HEP packages

Packages integrated into experiments are responsibility of experiment

IAIFI is focused much more on the development of concepts and ideas
The work in deployment is largely part of the experiment

29

A natural question: is this sustainable?

IAIFI is currently in its renewal cycle—decisions expected soon. The model is proving robust. Partner institutions have integrated IAIFI researchers into larger departments, embedding them in computer science and physics alike. Integration with CS departments remains challenging in academic structures, but that's a universal problem, not specific to IAIFI.

On the software side, IAIFI follows standard computer science practice: code released via GitHub and HuggingFace with datasets and benchmarks. Validation uses conventional ML metrics. Critically, because IAIFI operates at the algorithm design level rather than deep in HEP-specific software stacks, they have limited dependence on traditional HEP packages. This flexibility is a feature.

For code deployed in experiments—that becomes the experiment's responsibility. IAIFI focuses on algorithm innovation; experiments handle integration. This division of labor actually works remarkably well.

30 of 70

30

Colloquia

Seminars

Follow IAIFI

http://mailman.mit.edu/mailman/listinfo/iaifi-news

@iaifi_news

https://www.linkedin.com/company/iaifi/

Connect

IAIFI Affiliates:

Senior researchers/faculty in the Boston area interested in the IAIFI mission

https://iaifi.org/affiliates

Friends of IAIFI:

Junior researchers/students in the Boston area interested in the IAIFI mission

https://iaifi.org/junior-interest

Join mailing list

Follow on X (formerly Twitter)

Follow on LinkedIn

Upcoming IAIFI Public Colloquia

Get Involved with IAIFI!

(2:00–3:00 pm ET, in Kolker Room and on Zoom, open to MIT community)

Konstantin Rusch

Assistant Professor, Max Planck Institute for Intelligent Systems

Friday, November 21, 2025

Joint with CSAIL

Mathis Gerdes

Incoming IAIFI Fellow

Date TBA

Roger Melko

Professor, University of Waterloo

February 13, 2026

Lisa Everett

Professor, University of Wisconsin - Madison

Friday, February 27,2026

Roberto Trotta

Professor, International School for Advanced Studies (SISSA)

Friday, March 13, 2026

Tommaso Dorigo

Researcher, Italian Institute for Nuclear Physics (INFN)

Friday, April 10, 2026

…and more! Visit https://iaifi.org.events

30

31 of 70

680 awards with total 86M GPU/Node Hours awardeed

2 awards with 135K for particle and high energy physics
NAIRR240103 Uncertainty Quantification and Anomaly Detection with Evidential Deep Learning
NAIRR240337 Machine-learned particle-flow reconstruction

The National Artificial Intelligence Research Resource (NAIRR) will provide a shared national research infrastructure to bridge this gap by connecting U.S. researchers and educators to AI resources — computation, data, software, models, training and educational materials — to advance research, discovery and innovation. directed by Winning the Race: America's AI Action Plan,

Pilot award

32 of 70

AI/ML for HEP�Challenge and Opportunities

Growing data complexity and advanced detectors
Need for real-time workflows
AI for real-time event selection

Standardize detector data for AI
DAQ/calibration integration
Open science and interoperability

33 of 70

NSF Expects to make one (1) award at up to $35M for a period of up to 5 years.

Operations Center for all US AI research
User portal, AI-ready datasets, models
Emphasis on outreach and user support

Opportunities for HEP instrumentation

Compute: Advanced AI platforms for instrument development.
Data: Real-time analysis of experiment data streams.
Models: AI-driven simulation and self-driving controls.
Collaboration: National, multi-disciplinary research network.

34 of 70

NSF News

NSF and NVIDIA partnership enables Ai2 to develop fully open AI models to fuel U.S. scientific innovation

August 14, 2025

$152M NSF & NVIDIA partnership with Ai2 builds open AI models for U.S. science.
Multimodal AI tools accelerate research, code, and discovery.
Project supports national AI workforce and collaboration.

Ideas for HEP Instrumentation

Automate detector design and monitoring with AI models.
Enable real-time fault and data quality detection using AI.
Speed up firmware and analysis code creation with generative AI.

35 of 70

DOE AI/ML at HEP

Mission: Enable High Energy Physics discovery science through applications of Artificial Intelligence (AI) and Machine Learning(ML), and further understanding of fundamental AI/ML techniques
April 2024 – PCAST Report “The cosmologists and particle physicists … are some of the earliest adopters—and developers—of AI, so an epoch of advanced AI is an epoch of exciting discoveries in fundamental physics and cosmology.”

Programmatic AI/ML – where the objective of the research is to resolve existing technical challenges using ML

Core AI/ML – development that can realize the potential of AI/ML to benefit HEP’s mission and improve AI/ML methods. Leverage technical development supported beyond DOE HEP

Jeremy love DPF 2025

36 of 70

Core AI/ML Delivers New Capabilities

FAIR Universe – integrated codabench with NERSC to enable uncertainty aware ML challenges

Curated HEP datasets for LHC Higgs analysis and Cosmology Weak Lensing corrections

HiggsML increase sensitivity to Higgs to tau tau and reduce systematic uncertainties
Map N-body simulations onto fundamental parameters of interest
Framework being used outside of HEP

37 of 70

Hardware-Aware AI Awards

To be announced publicly soon
Advanced R&D into real-time low power edge AI and control systems in FPGAs and custom ASICs and enabling technologies

Maurice Garcia-Sciveres

Network intelligence for fault tolerance

38 of 70

How to Engage & Compete

Join teams with labs, universities, industry
Subawards and cross-disciplinary applications
Workforce development required in proposals

TAC-HEP WATCHEP C2-THE-P2 LGT4HEP

Keisuke Yoshihara: Modern Electronics Education with AI/ML and FPGA

39 of 70

2026 CPAD Workshop

Date: Oct 20-23, 2026

Venue: University of Washington

Co-hosted by UW and PNNL

Quentin Buat �

Isaac Arnquist��Maria Laura di Vacri

40 of 70

Summary

41 of 70

41

700PB/year

The Square Kilometre Array (SKA)

HL-LHC

100PB/year~EB/year

Next-Generation Big Data Frontiers

700PB/year

200PB/year

Advanced Photon Source (APS)

Linac Coherent Light Source II �(LCLS-II)

100PB~EB/year

Exabyte-scale Big Data Challenge

42 of 70

LHC leading Big Data challenge

42

Data volume, Streaming rates and Dimensionality of Data surpassed industry standard

2021

Data complexity across disciplines: particles, neurons, cosmic phenomena (gravitational waves).
AI algorithms combined with novel processors (GPUs, FPGAs) create new avenues for fast, scalable data analysis.
AI allows integrated study across scales impossible before.

As scientific data sets become progressively larger, algorithms to process this data quickly become more complex. In response, AI has emerged as a solution to efficiently analyze these massive data sets. Emerging processor technologies, such as graphics processing units (GPUs) and field-programmable gate arrays (FPGAs), allow AI algorithms to be greatly accelerated. The combination of AI and these processors is leading to a revolution in the way we analyze data, minimizing the time needed to perform the most advanced of analyses, and allowing us to address the challenges brought about by the omnipresent onslaught of data. To harness these new developments and leverage them for the advancement of science, the newly-created $15 million NSF HDR Institute of Accelerated AI Algorithms for Data-Driven Discovery (A3D3) aims to incorporate AI algorithms with new processors such that they can analyze the unprecedented data sets we face.

43 of 70

Critical challenges across multiple disciplines

43

low latency
high throughput processing,
real-time control modules, and
custom processing elements.

2025

Data complexity across disciplines: particles, neurons, cosmic phenomena (gravitational waves).
AI algorithms combined with novel processors (GPUs, FPGAs) create new avenues for fast, scalable data analysis.
AI allows integrated study across scales impossible before.

As scientific data sets become progressively larger, algorithms to process this data quickly become more complex. In response, AI has emerged as a solution to efficiently analyze these massive data sets. Emerging processor technologies, such as graphics processing units (GPUs) and field-programmable gate arrays (FPGAs), allow AI algorithms to be greatly accelerated. The combination of AI and these processors is leading to a revolution in the way we analyze data, minimizing the time needed to perform the most advanced of analyses, and allowing us to address the challenges brought about by the omnipresent onslaught of data. To harness these new developments and leverage them for the advancement of science, the newly-created $15 million NSF HDR Institute of Accelerated AI Algorithms for Data-Driven Discovery (A3D3) aims to incorporate AI algorithms with new processors such that they can analyze the unprecedented data sets we face.

44 of 70

Training

Inference

Very large datasets & memory, CPU/GPU/TPU farms, floating point required
Done in an AI/ML Framework (Tensorflow, PyTorch, etc.).

44

Credit: Marzieh Vaez Torshizi

Requires minimal computational resources
Often has real-time performance/power requirements that require custom hardware, e.g. FPGA/ASIC/Edge devices

Unique in HEP instrum.

45 of 70

Trending in industry

45

46 of 70

Challenge and Opportunities

Growing data complexity and advanced detectors
Need for real-time workflows
AI for real-time event selection

Standardize detector data for AI
DAQ/calibration integration
Open science and interoperability

47 of 70

Office of Science Initiatives

SC Research Initiatives are advanced technology initiatives, important to SC and to US science and tech national enterprise.
HEP encourages research proposals aligned with exiting initiatives

https://science.osti.gov/Initiatives

48 of 70

DOE AI/ML at HEP

Mission: Enable High Energy Physics discovery science through applications of Artificial Intelligence (AI) and Machine Learning(ML), and further understanding of fundamental AI/ML techniques
April 2024 – PCAST Report “The cosmologists and particle physicists … are some of the earliest adopters—and developers—of AI, so an epoch of advanced AI is an epoch of exciting discoveries in fundamental physics and cosmology.”

Programmatic AI/ML – where the objective of the research is to resolve existing technical challenges using ML

Core AI/ML – development that can realize the potential of AI/ML to benefit HEP’s mission and improve AI/ML methods. Leverage technical development supported beyond DOE HEP

Jeremy love DPF 2025

49 of 70

Programmatic AI/ML Strengthens �HEP Science

Recent discovery by DESI that Dark Energy is not static but evolves over time

Used ML to combine results from multiple experiments and for the first time determine in a model agnostic way, the Dark Energy behavior across the age of the universe (z)

This discovery received wide attention in global popular press with more than 1,500 articles in 35 languages
Use of AI/ML methods to process images and identify features

50 of 70

Core AI/ML Delivers New Capabilities

FAIR Universe – integrated codabench with NERSC to enable uncertainty aware ML challenges

Curated HEP datasets for LHC Higgs analysis and Cosmology Weak Lensing corrections

HiggsML increase sensitivity to Higgs to tau tau and reduce systematic uncertainties
Map N-body simulations onto fundamental parameters of interest
Framework being used outside of HEP

51 of 70

How to Engage & Compete

Join teams with labs, universities, industry
Subawards and cross-disciplinary applications
Workforce development required in proposals

TAC-HEP WATCHEP C2-THE-P2 LGT4HEP

Keisuke Yoshihara: Modern Electronics Education with AI/ML and FPGA

52 of 70

Backup

53 of 70

NSF Institutes for HEP

https://iris-hep.org/

https://iaifi.org/

https://a3d3.ai/

Three major NSF-funded institutes, each addressing unique aspects of computational and AI-driven research at the intersection of physics and data science.

aims to meet the software and computing challenges posed by the High Luminosity LHC (HL-LHC), developing state-of-the-art cyberinfrastructure and acting as a community-wide hub for software R&D in high energy physics.

focused on fusing foundational physics principles with cutting-edge AI approaches to tackle challenging problems in physics and galvanize innovation in trustworthy AI.

targets real-time AI solutions for large, complex datasets across high energy physics, multi-messenger astrophysics, and systems neuroscience, integrating customized AI with advanced hardware acceleration.

54 of 70

Jan-Frederik Schulte

Miaoyuan Liu

55 of 70

680 awards with total 86M GPU/Node Hours awardeed

2 awards with 135K for particle and high energy physics
NAIRR240103 Uncertainty Quantification and Anomaly Detection with Evidential Deep Learning
NAIRR240337 Machine-learned particle-flow reconstruction

The National Artificial Intelligence Research Resource (NAIRR) will provide a shared national research infrastructure to bridge this gap by connecting U.S. researchers and educators to AI resources — computation, data, software, models, training and educational materials — to advance research, discovery and innovation. directed by Winning the Race: America's AI Action Plan,

Pilot award

56 of 70

NSF Expects to make one (1) award at up to $35M for a period of up to 5 years.

Operations Center for all US AI research
User portal, AI-ready datasets, models
Emphasis on outreach and user support

Opportunities for HEP instrumentation

Compute: Advanced AI platforms for instrument development.
Data: Real-time analysis of experiment data streams.
Models: AI-driven simulation and self-driving controls.
Collaboration: National, multi-disciplinary research network.

57 of 70

NSF News

NSF and NVIDIA partnership enables Ai2 to develop fully open AI models to fuel U.S. scientific innovation

August 14, 2025

$152M NSF & NVIDIA partnership with Ai2 builds open AI models for U.S. science.
Multimodal AI tools accelerate research, code, and discovery.
Project supports national AI workforce and collaboration.

Ideas for HEP Instrumentation

Automate detector design and monitoring with AI models.
Enable real-time fault and data quality detection using AI.
Speed up firmware and analysis code creation with generative AI.

58 of 70

Current: AI on HEP Workflow

58

Digitize Data �Signal Processing

Raw Data archiving�Data formatting

Build New Algorithms �Develop AI Models �AI Inference

AI Inference

Sensor

Storage

Analysis

Data Reduction

Reduction

Meeting Big Data Challenges by Embedding AI Inference in Data Reduction for Efficient Storage and Real-Time Analysis

On-Line

Off-Line

Dickinson

Arghya Ranjan Das

4D tracking Serge Oktyabrsky, Timon Heim

59 of 70

Trend: AI-Enhanced Data Processing

59

Digitize Data �Signal Processing

AI Inference

Sensor

Data archiving�Discovery archiving

Storage

Build New Algorithms �Develop AI Models �AI Inference

Analysis

Data Reduction

Reduction

Signals a shift toward fully automated, intelligent scientific workflows driven by AI — from sensor data acquisition to storage.

On-Line

Off-Line

60 of 70

Emerging HEP Opportunities

Automated trigger/event selection
Self-optimizing calibration for sensors
Digital twins for detector operation
Model/standards teams improve reproducibility

61 of 70

Takeaway Messages

Instrumentation at the heart of federal AI investment

foundational AI on edge devices
AI-driven DAQ

Funding, infrastructure, and collaboration available
Workforce & training components required
Engage now to help shape the future

62 of 70

Backup

63 of 70

Further Information & Q&A

OE: science.osti.gov
NSF: nsf.gov/funding/opportunities/national-artificial-intelligence-research-institutes
NAIRR-OC Webinar online
Q&A

64 of 70

AI to Accelerate Science and Engineering Discovery Workshop 2023

Report

Identify key challenges, opportunities, and research priorities for integrating advanced AI/ML and data analytics into scientific and engineering domains over the next five years

65 of 70

Two out of 30 awards to particle physics

ACED 2435808: Hardware-Accelerated Graph Neural Networks for Real-Time Decision-Making in High Energy Particle Physics

develop FPGA-accelerated graph neural networks (GNNs) for the CMS High-Granularity Calorimeter (HGCAL)

ACED 2435957: ACED: Physics-informed Geometric Deep Learning for Astrophysical Neutrino Reconstruction in IceCube DeepCore

NSF24-541

66 of 70

67 of 70

Roadmap

DOE Transformational AI Models Consortium
NSF AI Research Institutes & NAIRR-OC
Focus on HEP instrumentation
How to participate

68 of 70

NSF National AI Research Institutes

Multi-year research & workforce hubs
2025 themes fit instrumentation
Cross-domain: AI + materials + sensors

69 of 70

DOE – Transformational AI Models Consortium

$30M, 2-year program
Domain-specific “self-improving” AI
Multi-institution collaboration encouraged

70 of 70

Neuroscience Demo - work in progress

70

Real time detection of neural states from high-density electrophysiology to drive closed-loop manipulations

Orsborn group

Algorithms by Shlizerman group