Invited Speakers

Scale to See: From Extreme-Scale Computing and Scalable AI to New Scientific Discovery

Talk abstract: Modern scientific problems increasingly demand more than faster computation—they require fundamentally new computational capabilities. This talk explores how the convergence of AI, mathematical modeling, and extreme-scale computing can transform scientific computing from solving known problems faster to enabling questions that were previously computationally inaccessible.Through examples in computational imaging, inverse problems, and Earth-system science, I will discuss how scaling AI to unprecedented model sizes, spatial and temporal resolutions, and ensemble sizes can reveal structure, quantify uncertainty, and enable scientific inference that is difficult or impossible at conventional scales. I will highlight how advances in AI algorithms, architectures, and HPC systems work together to create these new capabilities, and discuss opportunities and challenges for the next generation of scientific discovery. 

Exploring the Exascale Universe 

Talk abstract: Recent exascale cosmology simulations are producing multi-petabyte datasets with unprecedented volume, resolution, and physical fidelity in support of next-generation observational surveys. These datasets allow us to visualize the evolution of cosmic structure across enormous spatial and temporal scales and to generate synthetic observations for direct comparison with survey data. They also provide powerful inputs for AI, including emulators and digital twins, as well as agent-based approaches that can streamline complex, multi-step scientific queries and analyses across datasets whose scale and complexity make traditional exploration increasingly difficult. 

Designing Scalable Vision Models for Scientific Understanding 

Talk abstract: Scientific images are observations of an underlying physical or biological system. Learning from them requires preserving structures and relationships that matter scientifically, often across resolutions that make conventional vision models computationally impractical. This talk examines where we can intervene in vision models to address these challenges, connecting decisions about input representation and compute allocation to the supervision that shapes learned features. Drawing on our work in scientific imaging, we discuss adaptive patching for allocating computation and MICROFIELD, which uses microstructure descriptor fields to guide learning toward scientifically meaningful structure without additional inference cost. These examples illustrate how targeted changes to model design and training can support both computational efficiency and scientific fidelity. 

Faster Scientific Multi-dimensional data analytics with NVIDIA cuPhoton 

Talk abstract: Observatories and telescopes, lasers and X-ray light sources, and other high-throughput instruments generate image data faster than CPU-bound pipelines can process it to support timely decisions. The computational bottleneck is rarely one slow kernel, but rather an end-to-end workflow that spans from the sensor/simulation to a decision: load, read, HPC processing and analysis, AI-signal detection. Modern facilities routinely accumulate petabytes of multidimensional data across a single survey or experiment campaign. With CPU-first implementations, it can take hours to months—sometimes a year or more—to obtain scientifically useful results from data the instrument produced in seconds. These delays make it harder for scientists and engineers to extract insights to make informed decision and advance their research. What is missing is a GPU-native path that covers every stage from the sensor to the decision, not a faster version of any one step. I will talk about how NVIDIA cuPhoton provide open-source NVIDIA CUDA-X toolkit of GPU-accelerated building blocks for that path, spanning use cases from spectral and optical astronomy to time-domain laser/X-ray analysis. By keeping image data on the GPU from sensor read through classification—on either single to hundreds of GPUs—cuPhoton collapses the wait to as low as microseconds, resulting in a tight loop that researchers can run interactively, shortening time to insight and letting facilities keep pace with the data their instruments are producing.