Confab26 IP Demos

Demo Descriptions

Demo A+D: AI-driven Real-time Structure Identification from X-ray Spectroscopy

Plenary Demo: Deyu Lu, Scientist, Center for Functional Nanomaterials, and Stuart Wilkins, Department Chair, NSLS-II, Brookhaven National Laboratory

Breakout Demo: Deyu Lu, Sara Mason, Sairam Vatsavai, Dan Allen, Brookhaven National Laboratory

Description: This demo showcases LightshowAI developed at Brookhaven National Laboratory as an end-to-end, near-instrument ML-driven x-ray spectral analysis workflow, with clear extensibility to other user facility instruments. LightshowAI streamlines live x-ray spectral data from the NSLS-II beamline and uses AI to solve the atomic structure of measured spectra, providing physical insight to researchers for real-time knowledge extraction and enabling dynamical control in autonomous experimentation. 

Demo B: Real-time High-throughput Neural Compression for Particle Detector Data​​     

Plenary Demo: Yihui (Ray) Ren, Group Lead, CDS, Brookhaven National Laboratory

Breakout Demo: Ray Ren, Ai Kagawa, Brookhaven National Laboratory

Description: This demo presents a near-real-time, AI-driven data compression pipeline for the sPHENIX experiment at Brookhaven National Laboratory, developed under the AmSC project. Targeting future detector sparse output rates of 10–100 GB/s, the system employs a neural compression model (BCAE-VS) to enable streaming compression, reconstruction, and evaluation with a live feedback loop. The workflow integrates ML training and inference tracking, along with interactive visualization, to support scalable data handling for next-generation particle physics experiments. 

Demo C: Federated Orchestration of Heterogeneous Workflows​

Plenary Demo: Alexei Klimentov, Scientific Computing and Data Facilities Director, CDS, Brookhaven National Laboratory

Breakout Demo: Alexei Klimentov, Ofer Rind, Oszkar Tarjan, Horonori Ito, Matt Cowan, Brookhaven National Laboratory

Description: This demo showcases the orchestration of diverse scientific workflows across on-premises computing resources, commercial offerings and HPCs. It moves away from the 'facility silo' model. It also presents a new tool for translating physics inputs into production and automated tasks, including the LLM-based WMSbot. The LLM is equipped with MCP tools for accessing submitted job and task data and metadata. Integration with AmSC projects and partnerships is also discussed. 

Demo E1: AI-Assisted Algorithm Development and Tuning

Plenary Demo Only: Robert Edwards/JLAB, Michael Wagman/FNAL

Description: A human developer and a commercial coding agent work together to improve a neutrino-nucleus scattering simulation pipeline under AmSC that requires processing a petabyte of quark field inputs from Frontier.

Demo E2: Agentic Workflows for Lattice QCD

Plenary Demo: Robert Edwards/JLAB, Michael Wagman/FNAL

Breakout Demo: Chris Kelly/BNL

Description: Agentic AI workflows, including optimization and visualization agents, are being deployed to accelerate and simplify complex Lattice QCD calculations and research within the AmSC infrastructure.

Demo E3: AI Powered QCD Global Analysis

Plenary Demo: Robert Edwards/JLAB, Michael Wagman/FNAL

Breakout Demo: Nobuo Sato/JLab

Description: The JAM app under FemtoMind is connecting scientific expertise, AmSC infrastructure, and AI-guided flows, providing an AI-native ecosystem for next-generation scientific discovery.

Demos F1-F5

Plenary Demos Introduction: AmSC Scientific User Facilities Infrastructure Partnership (SUF IP) Project — Paolo Calafiura, Data & Computational Scientist, Scientific Data, LBNL

Plenary Demos Conclusion: Nicholas Schwarz, Principal Computer Scientist, ANL

Demo F1 Agentic Accelerator Controls

Plenary Demo: Remi Lehe, Research Scientist and Physicist, Advanced Modeling Program, Accelerator Technology & Applied Physics Division, LBNL 

Breakout Demo: Remi Lehe (LBL), Auralee Edelen (SLAC), Sara Miskovich (SLAC), Thorsten Hellert (LBL)

Description: We will demonstrate AI/ML systems for real-time guidance during the operation of particle accelerators, applied to the BELLA/LBNL and LCLS/SLAC facilities. These systems are exposed via an interactive dashboard and chat interface, and include facility-specific digital twins and AI agents.

Demo F2 Cross-Facility Data Ingestion and AI workflows for Nuclear Physics

Plenary Demo: David Lawrence, Scientific Software Scientist, JLAB        

Breakout Demo: David Lawrence (JLAB), Meifeng Lin (BNL), Armen Kasparian (JLAB), and Mohammad Atif (BNL) 

Description: High-bandwidth transfer of data from Brookhaven National Laboratory to the Thomas Jefferson National Accelerator Facility from a nuclear physics experiment. The demonstration will show a web interface with the points of connection to the American Science Cloud highlighted in the configuration/settings. Dynamic monitoring of the data flow will be shown. This demo will illustrate exposing the internal archive of process variables of the CEBAF accelerator to external users via the American Science Cloud infrastructure. This system will allow authorized external users to access the data for studies relevant to development of new accelerator designs and operation. 

Demo F3 Inference as a Service for HEP and NP Data Analysis

Plenary Demo: Xiangyang Ju, Computing Systems Engineer, SciFata, LBNL, and

Meghna Bhattacharya, Applications Physicist, FNAL           

Breakout Demo: Xiangyang Ju (LBL), Meghna Bhattacharya (FNAL), Nhan Tran (FNAL), Julien Esseiva (LBL)

Description: We demonstrate that the AmSc cloud platform enables scalable, low-latency Machine Learning inference for HEP workflows, accelerating scientific productivity. Specifically, we will leverage computing capabilities offered by the AmSC platform to accelerate production-level workflows from different frontier experiments including the ATLAS, CMS, and DUNE experiments.  

Demo F4 AI-Enabled Science at the Light and Neutron Sources with the American Science Cloud (five subdemos)

Plenary Demo (Breakout Demos below): Steven Henke (ANL)

Description: Demonstration of how  AmSC services and resources have unlocked a set of impactful, high-priority AI-enabled light and neutron source data analysis and discovery workflows to help address data deluge and data complexity challenges bringing the facilities one step closer to realizing a ubiquitous AI computing fabric that accelerates scientific discovery for experimental user science.

• F4 – Tomography (Tim Dunn/SLAC), Dylan McReynolds/LBNL, David Abramov/LBNL, Hannah Parraga/ANL): SSRL, APS, and ALS have developed tomography analysis workflows that carry raw detector data from beamline acquisition through reconstruction, segmentation, and visualization on HPC systems. These workflows leverage the IRI API for job orchestration and data movement, and integrate AmSC services including the Data Catalog and Model Catalog, where SAM3 models are registered via MLflow for AI-enabled segmentation. This demo will walk through the end-to-end pipeline, from raw data capture to interactive visualization of the segmented 3D volume.

• F4 - Ptychography (Steven Henke/ANL, Valerio Mariani/SLAC): We will demonstrate an AI/ML ptychography workflow between a light source (APS and LCLS-II) and AmSC compute facilities (ALCF, NERSC, and possibly OLCF). The workflow had three stages: (1) execute a conventional reconstruction algorithm to estimate the probe, (2) use the probe to train a physics-informed neural network, and (3) use the neural network for fast inference on an unseen dataset. All steps will be performed on remote resources using the AmSC Data Transfer Service for data movement and IRI API for compute job submission & monitoring. 

• F4 – Bragg Peak Detection (Valerio Mariani/SLAC), Jana Thayer/SLAC, Cong Wang/SLAC), Sam Welborn/SLAC): Data collected at the LCLS MFX beamline is transferred to NERSC. A mixed CPU/GPU starts on Perlmutter: the LCLS data reading frameworks reads and pre-processes the data and sends it to other processes that performs peak detection by running inference with the PeakNet algorithm. The peak information is saved in files that are transferred back to LCLS.

• F4 – Neutron Scattering (Sergey Yakubov/ORNL, Mathieu Docuet/ORNL): Web portal consuming live experimental data from SNS and running workflows on NERSC (Perlmutter) and Frontier or local cloud.

• F4 – Latent Space Exploration: Daniel Allan (BNL), Stuart Wilkins (BNL) Chat with an LLM to interrogate raw and processed data via Tiled, relying on AmSC MLflow, IRI API, and inference service. Integrate a model from collaborators at ALS to display live feedback during experiment, using AmSC MLflow, IRI  API, and inference service. Upload experiment metadata and links to data from Tiled promptly to AmSC OpenMetadata Catalog.

Demo F5 OpenCosmo on AmSC

Plenary and Breakout Demos: Patrick Wells, Computational Cosmologist and Software Engineer, ANL

Description: We will demonstrate using the platform both from a web portal and in an agentic environment and discuss the scientific impact the platform is already having. We will demonstrate a Globus action provider for the AmSC IRI API. We will discuss the challenges of developing applications in an HPC context, and the ways in which AmSC is presently and can in the future provide better tooling for solving these problems.

Demo G: HAIDIS: Hardware-enabled AI Distributed Inverse Problem Solver

Plenary and Breakout Demos: Ilya Baldin, Acting HPDF Technical Director, JLAB and Daniel Lersch,  Supervisor CST Data Scientist, JLAB

Description: HAIDIS team will demonstrate the two phases of the workflow — training on GlueX data and inference from the trained model snapshots. We will review the results obtained with HAIDIS in the context of the ongoing Gluex Dalitz Plot analysis. We will demonstrate the use of EJFAT load-balancing system and integration with AmSC via IRI compute API for training and with MLFlow for inference. Results will be visualized using custom dashboards and notebooks.

Demos H1-H3

Plenary Intro: The SLAC/S3DF Infrastructure Partner - Jay Srinivasan, Division Director, Scientific Computing Systems (SLAC)

Demo H1: Automated LCLS Run Processing Across S3DF and NERSC via the IRI Interface

Plenary Demo Only: Amith Murthy, Senior Software Developer, and Sam Welborn, Big Data Architect, SLAC

Description: Tackling SLAC's unique scientific computing challenges requires the ability to orchestrate resources across various DOE facilities. This talk will cover S3DF's efforts in building out the IRI API for SLAC and integrating it in our scientific workflows. The demo will showcase one such workflow at LCLS, which now runs at both NERSC and S3DF through IRI.

Demo H2: Scalable, Automated AmSC Catalog Registration through Tiled

Plenary Demo Only: Sam Welborn, Big Data Architect, SLAC

Description: This presentation highlights S3DF's active role in AmSC data services. The focus of this demo is our cross-DOE work on Tiled, an open-source data access software supported by a worldwide collaboration. We will showcase how recent hackathon-initiated code contributions to Tiled enable horizontally scalable, automated ingestion of metadata directly into the AmSC Data Catalog. We present how this feature is being used in a ModCon project (MAIQMag) to streamline discovery of physics artifacts to be used in multimodal AI training workflows.

Demo H3: SLAC S3AI — A Sandbox for Streaming Inference and Control

Plenary Demo Only: Ryan Coffee, Senior Scientist, SLAC

Description: The SLAC Sandbox for Streaming AI (S3AI) is one of SLAC’s contributions to the American Science Cloud. It serves as a collaboration space for tri-sector development — Labs, Industry, Academia — for real-time streaming AI inference at the microsecond level.  We will show two to three pre-recorded demonstrations, Machine Vision at 90k frames/second, fusion plasma shape inference at 10 microsecond latency (100k frames/second), and an industry partnership to test external ideas regarding secure multi-site data and model-weight sharing. 

Demo I: Stellar-AI, Fusion Digital Twin

Plenary Demo: Shantenu Jha, Computational Sciences Department Head, PPPL

Description: The STELLAR-AI demo will demonstrate the integration of heterogeneous AI and HPC services (AtScale & Model) ROSE & RHAPSODY, achieving state-of-the-art performance in plasma reconstruction and real-time state forecasting. This milestone integrates American Science Cloud (AmSC) capabilities, leveraging AtScale & Model Services and federated data services to bridge national laboratory facilities with scalable AI workflows. 

Demo J: Multi-agent Workflow for Both Nuclear Plant-level and Reactor Component Design

Plenary and Breakout Demos: Brandon Biggs, High Performance Computing Systems Administrator, INL

Description: Prometheus is a multi-agent orchestration platform purpose-built for the nuclear reactor lifecycle. This technical demo will focus on specialized agents that handle discrete workflow domains and the challenges that exist with long-horizon tasks. Integrations with American Science Cloud services will be highlighted and discussed. Partners include: Idaho National Laboratory, Sandia National Laboratory, Argonne National Laboratory, Oakridge National Laboratory, AmSC, Modcon, AWS, Microsoft, NVIDIA, Oklo, Westinghouse, Aalo, EverstarAI, Atomic Canyon.

Demo K: HERO capabilities

Plenary Demo: Kristi Potter, Data Visualization Scientist, NRL

Breakout Demo: Monte Lunacek, NLR

Description: HERO is the NLR framework for providing our lab researchers access to AmSC services as they become available. The most important capability that we want to contribute is distributed, asynchronous agentic tools that can execute long-running jobs on our infrastructure. This video will show some additional efforts to engage with AmSC.

Demo L: Showcasing Data and Services at scale - the EDX®-AmSC Ecosystem from User Interface to Service Integration

Plenary Demo Only: Chad Rowan, Research Scientist, Computational Science & Engineering, and Jack Sarle, Lead Engineer for AI & Software Systems, NETL

Description: The NETL EDX IP Team will showcase the Energy Data eXchange® (EDX) Ecosystem, including the core EDX and EDX Spatial applications and the EDX AI Incubator, a secure GCP researcher environment, as a suite of capabilities for potential integration into the AmSC platform. Additionally, we will demonstrate our ongoing data integration services, illustrating how updated ingestion code and each platform’s APIs can work together to accurately synchronize EDX metadata within the AmSC catalog. Finally, EDX is an Oauth2 provider, enabling additional log in support via DOE-OneID, and is poised to support integration within AmSC’s web applications for future agility of this ecosystem. 

Demo M: Fermi Data Platform — Uploading, Curating, and Accessing datasets

Plenary Demo: Oliver Gutsche, Senior Scientist, Scientific Computing, FNAL

Description: Fermi Data Platform will demo uploading a dataset to the platform, registering metadata of the dataset in the local metadata catalog, see the automatic sync register the dataset in the global AmSC metadata catalog, then a user will query the global catalog and use the AmSC Data Movement API to copy the data out of the Fermi Data Platform to maybe NERSC.

Demo N: Interactive DIII-D Digital Twin for Experimental Science 

Plenary and Breakout Demos: Sterling Smith, Mark Kostuk, and David Schissel, Magnetic Fusion Energy Division, General Atomics

Description: A live demonstration of the DIII-D Digital Twin in concert with ongoing operations of the DIII-D National Fusion Facility (schedule permitting). A full-pipeline example will be shown of how this new tool can be used in production to enhance the Facility's scientific capability. Incorporation with the AmSC infrastructure enables many predictive scenarios to be remotely executed at NERSC and ALCF using the At-Scale API under the time constraints of ongoing experiments. The Fusion Data Platform both provides data movement and is an infrastructure component being provided back to the AmSC. Many faster-than-realtime surrogate models have been trained on IRI-obtained datasets and tuned with the aid of MLFlows. These models enable the digital twin to be performant, and the platform of integration demonstrates a repeatable pattern for digital twins of other facilities.  A natural language interface enabled by the AmSC AI-Services allows our operators to interrogate the newly simulated data and select superior options for immediate use in the control room.