Understanding the recombination and ionization reactions in bicarbonate - carbonic acid equilibrium

Prabhat Prakash, Caltech

In my talk, I will discuss results from extensive simulations performed with new GPU nodes of Perlmutter which lead to an atomistic understanding of pH and effect of local structure of solvated bicarbonate ions, and lifetimes of various reaction species in water microenvironments.

Fast Computational Modeling of Gravitational Lenses using NERSC GPUs

Saul Baltasar, Nicolas Ratier-Werbin, Lawrence Berkeley National Lab

We apply a state-of-the-art GPU-accelerated modeling code to two observed Einstein Crosses - particularly striking gravitational lensing systems. We achieve excellent results. Both systems are modeled with record time: 2 orders of magnitude speed-up compared with other modeling codes, all CPU-based.

Self-Supervised Learning for Sparse Computational Imaging

Vidya Ganapati, Lawrence Berkeley National Laboratory

We present a data-driven, self-supervised approach to image reconstruction from sparse, noisy measurements. No ground truth is required, only a dataset of sparse measurements on different sources. LED array microscopy, computed tomography, and crystallography are used as illustrative examples.

HydraGNN and DDStore for Scalable Training of Graph Neural Networks on Millions of Organic Molecules

Jong Youl Choi, Oak Ridge National Laboratory

We introduce HydraGNN and DDStore for large-scale graph neural network training. We highlight the data challenges and the performance for running a challenging GNN training with 10 million molecules using Summit and Perlmutter. We share our experience and performance comparison.

Towards Unsupervised Anomaly Detection in Production HPC Systems

Efe Sencan, Boston University

This talk highlights our recent ML framework for computer system analytics for performance anomaly detection. We will be introducing our variational autoencoder based framework that detects performance anomalies on compute nodes using multivariate time series telemetry data.

Simulations of the Antarctic Ice Sheet from 2000–2300 using the MPAS-Albany Land Ice model

Trevor R. Hillebrand, Los Alamos National Laboratory

We present an ensemble of 300 year simulations of the Antarctic Ice Sheet under various climate forcings. We find that Antarctica is likely to contribute up to 3 m to global sea level by the year 2300, but uncertainties are large due to poorly constrained model physics and boundary conditions.

Anharmonic effects in hexagonal Zinc. Inelastic neutron scattering guided by ab initio and molecular dynamics simulations.

Vladimir Ladygin, California Institute of Technology

We present analyses of the experimental and computational results for hexagonal Zn. The diffuse features were observed in the experiment found to be consequence of phonon-phonon interaction. The experiment was supported with machine learning driven large scale simulations and ab initio calculations.

First Principles Investigation of Bulk Skyrmion Formation within the 50% Co-doped Fe5GeTe2 Layered Magnetic Metal

Tyler Reichanadter, UC Berkeley -- EFRC (NPQC)

The layered magnetic metal Fe5GeTe2 (F5GT) has remarkable magnetic behavior that is sensitive to Cobalt substitution, which at 50% hosts Neel-type Skyrmions. We perform DFT calculations over many Co-F5GT motifs to elucidate the coupling between geometry, chemical composition, and magnetic order.

Superionic properties of H-C-N-O compounds studied with ab initio and machine learning molecular dynamics simulations

Kyla Briann de Villa, UC Berkeley - Department of Earth and Planetary Science

Using ab initio and machine learning molecular dynamics simulations, we studied transport properties of 13 H-C-N-O materials at extreme pressures and temperatures, finding superionic phases, in which protons diffuse through stable lattices of heavier species, to be prevalent in this chemical space.

Petabyte-scale Image Processing Tools for Light Sheet Microscopy Data

Xiongtao Ruan, University of California, Berkeley

Light-sheet microscopy is a powerful technique for real-time 3D imaging of cellular and subcellular dynamics. Here, we developed the software LLSM5DTools, a scalable and efficient image processing pipeline to handle petabyte-scale datasets.

Scientific Workflows with Pegasus

Karan Vahi, USC Information Sciences Institute

The talk will give an introduction to Pegasus WMS and describe recent work done to enable NERSC users to submit workflows to Perlmutter. This new approach allows users to spin up a workflow submit node container in the Spin environment that is fully configured to submit Pegasus workflows to NERSC.

Advancing Flood Frequency Analysis through a Fully Distributed Integrated Hydrological Model and High-Performance Computing

Gabriel Perez, Oak Ridge National Laboratory

This study combines a physical-based hydrological model, storm transposition techniques, and high-performance computing to estimate floods from 10,000 extreme storms in Texas. Our findings empower advancements in flood frequency analysis and deepening our understanding of this natural phenomenon.

First principles studies of carbon capture chemisorbents
Jonathan Owens, General Electric Research

We will present our NERSC-enabled work to scale up computational characterization of carbon capture chemisorbents. We will in particular focus on first principles prediction of CO2 capacity and structural stability and our pipeline integration with Perlmutter-GPU.

Turbulent transport in LHD stellarator with boron impurities
Tajinder Singh, Indian Institute of Science, Bangalore, India

In this talk, a gyrokinetic analysis of one of the experimental discharges of impurity injection experiments in LHD is presented using gyrokinetic toroidal code (GTC), and a comparison of the simulations against the experimental observations is made.

Strong coupling of excitons and photons via ab initio calculations
Christopher Ciccarino, Stanford University

We develop a first-principle approach to study excitons-polaritons – excitations in materials where electrons and photons couple strongly. Our large-scale calculations reveal for the first time how their wavefunctions are strongly modified, providing new tunability for transport and quantum physics.

Proximity-induced moiré exciton in a transition metal dichalcogenide monolayer
Sudipta Kundu, Stanford University

We explore a noninvasive way to induce a spatially varying moiré potential in a monolayer transition metal dichalcogenide (TMD) externally and study the tunable novel excitons with spatial localization that differ from those possible in regular stackings of twisted bilayer TMDs.

SwiFT: Swin 4D fMRI Transformer
Junbeom Kwon, Seoul National University

Presenting our computation-efficient deep neural network capable of directly learning brain dynamics from 4D brain fMRI, we showcase the effective utilization of NERSC's resources and training opportunities, highlighting the substantial performance increase achieved through extreme-scale training.

A GPU-accelerated simulation and reconstruction chain for the DUNE near detector

Matt Kramer, LBNL

DUNE's liquid argon near detector will use a novel pixelated charge readout whose granular 3D images present interesting computational challenges. To meet them, the LBNL DUNE group has developed a detector simulation on Perlmutter GPUs and has integrated it into a NERSC-hosted full sim/reco chain.

Qubit Lattice Simulations of Electromagnetic Scattering from Tensor Dielectrics

George Vahala, William & Mary

A qubit lattice algorithm (QLA) is developed for the scattering of electromagnetic waves from tensor dielectric objects. QLA is ideally parallelized on classical supercomputers and encodable onto quantum computers. Detailed multiple reflected/transmitted wavefronts from dielectrics are seen.