An LLM-powered chatbot built on RAG over ALCF documentation answers HPC user queries about system usage, job scheduling, and software debugging with high accuracy.

AERIS is an AI-augmented earth system model trained on Aurora that delivers skillful climate predictions by combining physics-based simulation with large-scale deep learning.

A comprehensive architecture description of Aurora, Argonne's Intel-GPU exascale system, covering the hardware design, software stack, and early science results across multiple domains.

LLMs can reason over electrochemical literature to autonomously propose, evaluate, and prioritize battery material hypotheses, suggesting a new role for AI in accelerating energy storage research.

HiPerRAG achieves high-throughput scientific document retrieval by co-designing the retrieval pipeline with HPC storage, enabling exascale-grade RAG for large scientific corpora.

A survey of HDF5 advancements — async I/O, Cache VOL, subfiling, and new VOL plugins — that collectively deliver efficient, scalable parallel I/O for the exascale era.

Copper coordinates node-local storage caching across all nodes of an exascale job to eliminate redundant data loads and saturate GPU compute with training data.

We scale multimodal protein design beyond ExaFLOPS by combining Direct Preference Optimization with efficient data pipelines on the Frontier supercomputer.

DFTracer captures I/O and compute events across the full AI training stack with near-zero overhead, enabling detailed bottleneck diagnosis in complex multi-node workflows.

A physics-informed graph neural network ensemble detects higher-order gravitational wave modes from spinning black hole mergers, extending sensitivity beyond what standard matched filtering achieves.

h5bench provides a flexible, reproducible suite of HDF5 I/O micro-benchmarks covering a wide range of access patterns and VOL configurations for pre-exascale and exascale evaluation.

Fusing complementary electron tomography modalities yields 3D chemical maps at 1 nm resolution, opening new pathways for understanding catalysts, batteries, and nanostructured materials.

A graph AI ensemble incorporating physical priors achieves real-time gravitational wave detection competitive with Bayesian matched filtering at a fraction of the computational cost.

Real-time 3D tomographic reconstruction during live electron microscopy enables mid-experiment decision-making at near-atomic resolution, fundamentally changing how materials experiments are conducted.

A transparent HDF5 plugin stages writes and prefetches reads on node-local NVMe SSDs, delivering up to 10× I/O speedup for checkpoint-intensive applications with zero code changes.

Inference-optimized deep learning on HPC systems achieves real-time gravitational wave detection throughput, enabling continuous monitoring of LIGO data at production scale.

An interpretable deep learning model accurately predicts numerical relativity waveforms for spinning binary black holes, providing both high fidelity and physical insight into which features matter most.

Stimulus co-designs data staging and prefetching with the HPC storage hierarchy to eliminate I/O bottlenecks in multi-stage scientific AI pipelines.

DLIO faithfully reproduces the I/O patterns of scientific deep learning workloads, giving storage vendors and HPC centers a reproducible, configurable tool for measuring AI storage performance.

Dynamic compressed sensing adapts the measurement basis during a live electron tomography experiment, achieving accurate 3D reconstruction from far fewer projections than conventional methods.

A systematic OpenCL portability evaluation across Intel and NVIDIA GPU architectures reveals which benchmark categories achieve near-native performance and where vendor-specific tuning remains necessary.

A position-dependent hybrid functional adapts exchange mixing to the local dielectric environment in heterogeneous systems, achieving GW-level band-gap accuracy at DFT computational cost.

Deep learning galaxy classification scales to 2,048 GPUs on Blue Waters, enabling petabyte-scale Dark Energy Survey catalog construction without expensive human labeling.

Roofline analysis of leading HPC benchmarks across KNL and GPU architectures identifies which applications are memory-bound vs. compute-bound and guides optimization priorities.

Density matrix downfolding extracts effective low-energy model Hamiltonians directly from QMC wavefunctions, bridging ab initio many-body theory and model Hamiltonian physics.

QMC calculations reveal that σ-bonding electrons make a critical and previously underestimated contribution to electron correlation energy in graphene, with implications for low-energy effective theories.

Quantum Monte Carlo resolves the longstanding debate on the VO₂ metal-insulator transition, showing that strong correlation — not structural distortion alone — drives the phase change.

A systematic density-matrix downfolding framework extracts effective model Hamiltonians from ab initio many-body wavefunctions, providing a rigorous route from first-principles to correlated lattice models.

Dirac cones can be engineered in 2D photonic crystals through accidental degeneracy, enabling near-zero-refractive-index materials with cone-shaped dispersion analogous to graphene.

Engineered electromagnetic modes in a metallic cavity produce macroscopically sizable forces between plates, demonstrating a pathway to optomechanical actuation at microwave frequencies.

A unified framework shows that a metamaterial slab can continuously transition between perfect-lens imaging and electromagnetic cloaking by tuning its constitutive parameters.

Active exterior cloaking devices designed via boundary element methods can hide objects from external electromagnetic observation while the objects remain freely accessible, unlike passive shell cloaks.

Folded-geometry transformation optics provides a general strategy for relocating, multiplying, or illusion-converting electromagnetic sources using anisotropic metamaterial shells.