Sparsity, Equivariance, and Submicron-Resolved Geometry for Exascale Investigations

Thursday, February 6, 2025 3pm to 4pm

Special Applied Math Lecture

"Sparsity, Equivariance, and Submicron-Resolved Geometry for Exascale Investigations"
Diego Rossinelli, Adjunct Professor in the Institute for Computational and Mathematical Engineering (ICME) and Associate Director of the PSAAP III Center at Stanford University

Thursday, February 6
3 - 4pm
Maxwell Dworkin G125 

 

Synchrotron-based imaging is a primary data source for studying complex inorganic and biological structures, such as whole sections of the human optic nerve or whole murine kidneys, both in 3D and 4D. However, raw data alone does not immediately yield insights. Large-scale processing—ranging from signal enhancement and image segmentation to computational modeling—is essential to extract geometry, capture kinematics, and ultimately uncover dynamical properties.

 

A key challenge is devising computational schemes that integrate both machine-learned models and those grounded on first principles. Central to this endeavor is the design of equivariant discrete filters which respect the fundamental symmetries of the governing equations and are capable of identifying interactions across multiple spatiotemporal scales. Sparsity-promoting representations are main enablers of large-scale processing. The associated algorithms, however, often exhibit irregular access patterns and control-flow-rich instruction streams, which impair execution performance from a microarchitectural standpoint, potentially undermining the entire investigation by eccessively inflating time-to-solution on contemporary HPC systems. This challenge is further exacerbated when dealing with complex geometries, which often fail to meet the regularity assumptions made by many operator diagonalization schemes. 

 

In this presentation, I report on my development of volumetric multirate filter banks, starting with diverse wavelet families and extending to more advanced filtering methodologies, with emphasis on those achieving invariance or equivariance under translation, rotation, or scaling transformations. I will illustrate how these filters enable in-silico investigations of terabyte-scale images, which reveal detailed structural and flow-dynamics insights within complex geometries beyond what either imaging or CFD alone can achieve.

 

Speaker Bio: Diego Rossinelli is an Adjunct Professor in the Institute for Computational and Mathematical Engineering (ICME) and Associate Director of the PSAAP III Center at Stanford University. His research interests revolve around supercomputing investigations combining High-Resolution Imaging with CFD/FSI, and Biomedical Applications. In the PSAAP project, his current focus is on supercomputing, computational modeling, and visualization. Diego Rossinelli graduated from ETH Zurich in 2006, where he obtained the doctoral title in 2011. He received the ETH Medal, the Euro-Par Distinguished Paper Award, the ABB Research Prize, and the ETH nomination for the ACM Doctoral Dissertation Award. He was the recipient of the APS Milton van Dyke Award in 2012, the ACM Gordon Bell Prize in 2013, and finalist of the ACM Gordon Bell Prize in 2015. In 2016, Dr. Rossinelli left academia for a multi-year industry experience, and in 2018 was appointed as CTO of Lucid Concepts AG. In 2021, he returned to academia as a supercomputing specialist for Kantonsspital Aarau and the Faculty of Medicine at the University of Zurich.