Introductions from NERSC Leadership
Sudip Dosanjh, Jonathan Carter 

NERSC Overview and Year in Review
Rebecca Hartman-Baker

NERSC 10 Presentation
Presenter: Hai Ah Nam, NERSC

Coffee Break

NERSC 10 Panel Q+A Session
Panelists: TBA

Remote User Talks
Lunch provided only for registered, in-person attendees.

To Atomic Resolution and Beyond
Presenter: Stephanie Ribet, National Center for Electron Microscopy, Molecular Foundry, Lawrence Berkeley National Laboratory
Abstract: High-resolution electron microscopy (EM) plays a pivotal role in characterization at the nano- to atomic-scale across physical and biological sciences. Advances in hardware and software have changed the way that EM data is collected. Individual experiments now generate extremely large high-dimensional datasets in seconds. Integrating HPC into EM through acquisition, processing, and automation is essential for gaining novel insights into samples that are not possible with conventional methods.

Accelerating Lattice QCD Inversions with Multigrid and Deflation
Presenter: Leon Hostetler, Indiana University Bloomington
Abstract: Lattice QCD enables key precision calculations in the Standard Model but is computationally expensive due to repeated Dirac operator inversions. These are typically done with Krylov methods like conjugate gradient, which slow dramatically near the physical point. I present results from Perlmutter-GPU showing that eigenvector deflation and multigrid methods can accelerate these inversions, achieving up to 17x speedup for the most challenging cases.

Creating CGYRO ensembles with XGYRO to maximize fusion science output
Presenter: Igor Sfiligoi, University of California San Diego
Abstract: We recently realized that the largest memory user in the CGYRO fusion simulation tool is the constant collisional tensor that often does not change between simulation runs in many fusion studies. We thus recently developed a new tool, called XGYRO, that runs many such simulations as an ensemble in one MPI job, allowing for sharing that constant buffer. This allows for substantial reduction of communication overhead. We use XGYRO to study variation of ExB shear, relevant to H-mode in tokamaks.

AlphaSAXS: Reconstructing Protein Structure with Physiologically Relevant Conformations from Small Angle X-ray Scattering Data
Presenter: Feng Yu, LBNL
Abstract: AlphaFold revolutionized structural biology but struggles with flexible protein regions critical in Alzheimer's and COVID-19. To address this, we integrated small-angle X-ray scattering (SAXS) data to retrain the AlphaSAXS model with improved inference process, enhancing predictions for flexible regions. This approach advances experimentally guided structural prediction and improves computational modeling of physiologically relevant conformations.

Applying supercomputers to high precision analytic calculations of gravitational waves
Presenter: Zvi Bern, UCLA
Abstract: Future gravitational wave observatories present a highly nontrivial challenge to theorists to make predictions that match the expected experimental precision.  I will explain how we utilize supercomputers to advance the state of the art in analytics calculations, employing methods derived from quantum field theory, which are also crucial in precision collider physics cross-section computations.

Coffee Break

Bayesian Inversion Powered Digital Twin for Tsunami Early Warning on the Cascadia Subduction Zone
Presenter: Sreeram Venkat, Oden Institute, UT Austin
Abstract: We present a Bayesian inversion-based digital twin that uses acoustic pressure data from seafloor sensors, along with 3D coupled acoustic–gravity wave equations, to infer spatiotemporal seafloor motion in real time and forecast tsunami propagation for early warning with quantified uncertainties. We show results for this digital twin on the Cascadia subduction zone, where we exactly solve 1 billion parameter Bayesian inverse and prediction problems in 0.2 seconds using 512 GPUs on Perlmutter.

Containerization at NERSC Perlmutter: Performance and Security Challenges in HPC Scientific Workflows
Presenter: Vijayalakshmi Saravanan, University of Texas at Tyler
Abstract: We present a Singularity-based container workflow on NERSC’s Perlmutter (supported by DOE ASCR FAIR Award DE-SC0024492) for real-time analysis of large-scale molecular dynamics simulations. Using a producer-consumer architecture with ADIOS and Kernel CUSUM, our system enables efficient streaming, adaptive sampling, and change point detection. We address challenges in performance, security, and scalability, demonstrating improved resource utilization across HPC environments.

Active learning of MLIP accelerated by NERSC
Presenter: Beihan Chen, Pennsylvania State University
Abstract: Our project aims to understand the interplay between hydrogen and chemical short-range order (CSRO) in multiple-principal-element alloys. To investigate hydrogen's impact, a machine learning interatomic potential (MLIP) is trained using an active learning process, iteratively enhancing the training dataset with preliminary MLIP versions. A workflow managing NERSC jobs with varying QOS policies accelerates this active learning, balancing cost and efficiency.

Type Ia supernovae from Helium ignited double detonations in White Dwarf mergers
Presenter: Khanak Barghava, Stony Brook University
Abstract: Type Ia supernovae are powerful tools in cosmology and likely arise from exploding white dwarfs. A leading cause of this explosion is the merger of two white dwarfs, also known as the double degenerate channel. If both have thin helium shells, a surface helium detonation might trigger a core carbon detonation. My research uses simulations from AMReX-Castro to study the conditions under which this secondary detonation occurs, helping us better understand these stellar explosions.

NUGEx Welcome

Presenter: Emily Belli, General Atomics

Integrated Research Infrastructure (IRI) Updates
Presenter: Debbie Bard, NERSC

Coffee Break

DUNE at NERSC: Past, present, and future
Presenter: Matt Kramer, LBNL
Abstract: DUNE will use the world's most powerful neutrino beam to probe the mysteries of the Universe. NERSC has long played a key role in DUNE, supporting design studies, simulations, and mock analyses, as well as, more recently, prototype detector operations. As DUNE's full-scale detectors come online, they will present rich opportunities to use Doudna's advanced capabilities. This talk will present recent achievements enabled by NERSC, as well as the many interesting challenges that lie ahead.

gpAMR: Scalable AI/ML-Control of PDE-Based Simulation
Presenter: Chris Paolini,  San Diego State University
Abstract: gpAMR is an integrated SciML and ModSim capability for outer loop control of inner loop forward simulation that combines Chombo-based simulation (inner loop) with Gaussian process, functional analysis in gpCAM (outer loop). gpAMR was initially developed for stochastic decision making in traditional adaptive mesh refinement but in general provides UQ-control in situ simulation with real time decision making. We demonstrate scalable computations on Perlmutter GPUs for turbulent flow.

ML Surrogates for Ocean Dynamics Simulation
Presenter: Ahmed Burak Gulhan,  Argonne National Laboratory
Abstract: We leverage NERSC Perlmutter's GPU resources to develop ML surrogates for MPAS-Ocean simulations to advance ocean modeling & accelerate climate science. We generated our training data using MPAS-Ocean for multi-year simulations with different parameter configurations. Using PyTorch on Perlmutter (up to 16x4 GPUs), we explore different physics informed losses & neural networks along with training frameworks to efficiently and accurately model complex ocean system dynamics & quantify uncertainty.

Neural Decoding of Dexterous Hand Control using Perlmutter
Presenter: William Liberti, LBNL
Abstract: We are developing a new non-invasive biosensing approach to continuously decode dexterous hand movements from underlying neuromuscular control signals in real-world conditions. Using Perlmutter, we trained deep convolutional networks on 100+ hours of data from 20+ individuals performing semi-structured tasks. These models can generalize across users and environments, paving the way for practical high-bandwidth human-computer interfacing, and imparting human-level dexterity onto robotic systems.

The Real-Time Data Workflow of LZ Dark Matter Experiment at NERSC
Presenter: Maris Arthurs, SLAC
Abstract: The LUX-ZEPLIN (LZ) experiment is a leading direct detection search for dark matter, utilizing a dual-phase xenon time projection. LZ has set world-leading limits on WIMP-nucleon interactions, thanks in part to its robust hardware and sophisticated software infrastructure. This talk will provide an overview of the LZ offline software ecosystem, with a focus on the automated data movement and  processing performed on the Perlmutter supercomputer. 

Closing Remarks and Adjourn