Welcome to the group page of the Interdisciplinary Scientific Computing Laboratory (ISCL)!. ISCL performs research at the intersection of data science, applied mathematics, and high-performance computing to enhance the understanding of complex multifidelity and multiphysics phenomena in various applications. In other words - we create science-based AI algorithms for applications such as weather and climate modeling, disruption mitigation in nuclear fusion, data and model fusion for complex fluid flows, surrogate models for chaotic dynamical systems, and more.

ISCL is housed in the College of Information Sciences and Technology at Pennsylvania State University, as well as the Mathematics and Computer Science Division at Argonne National Laboratory. A high level overview of our research may be found in the following talks: [1], [2], [3], [4], [5]. Further information about publications can be found on Google Scholar and our software contributions are available on Github. ISCL has access to multiple HPC resources such as Bebop/Swing/Polaris (at Argonne), Roar (Penn State), and Perlmutter (NERSC). ISCL eagerly welcomes possibilities for education, collaboration, and consulting! Feel free to reach out to us for any questions.

News

Excited to present an invited talk at WPI's Param-Intelligence Seminar Series for scientific machine learning! Thank you Prof. Ameya Jagtap for the invitation!
Our paper on multipoint penalty optimization for training neural ODEs for chaotic systems is published in CMAME! Congratulations DJ for leading this effort! Read our preprint here.
Announcing a new preprint for non-Gaussian data assimilation of nonlinear observations led by Zachariah Malik at UC Boulder. A transformation of variables to a latent space enables more accurate application of conditional-Gaussian based Kalman filtering. Read our preprint here.
New preprint related to the use of distributed graph neural networks for spectral super-resolution led by Shivam! Applied for subgrid scale reconstruction of the Taylor-Green Vortex case at high Reynolds numbers. Read more here.
ISCL is pleased to announce two new projects! The Army Research Office has awarded us a prestigious Early Career Program Award (ECP) for studying connections between data assimilation, differentiable physics, and closure modeling. The Office of Atmospheric Sciences Research at the Department of Energy has also supported a collaboration with Prof. Matt Kumjian at Penn State Meteorology to use modal decomposition techniques for pattern analysis of aerosol and cloud interactions over Texas. Congratulations to all team members!
DJ, Haiwen, and Zach present posters at UQ-MLIP with DJ winning the best student poster award! Congratulations to all of them for representing ISCL so brilliantly. Romit also presents a talk on GNNs for SciML with work led by Shivam and Hojin.
See other archived news here.

Research information

The research manifesto of the interdisciplinary scientific computing group.

Grand challenge problems

Improving geophysical forecast models with data science. We are building models to forecast variables using scientific machine learning. (credit Troy Arcomano - collaborator).

Tokamak disruption mitigation for nuclear fusion (image taken from Boozer et al., 2012). We are building closures and surrogate models to accelerate the computation of complex nuclear fusion reactors (i.e., system level simulations).

Scientific ML algorithm development

An algorithm that builds deep learning function approximation for sparse, unstructured, and time-varying observations (Nature Machine Intelligence 3 (11), 945-951, 2021).

A novel graph neural network architecture that not only makes predictions but also identifies salient physical features in an interpretable latent space (credit Shivam Barwey - postdoc).

Applied machine learning

Building a wind-turbine wake model for on-shore wind farms in the Texas panhandle using LIDAR and meteorological data (Neural Computing and Applications 34 (8), 6171-6186, 2022).

Characterizing the schooling of fish using generative machine learning. Here, the variation in the density of fish is captured by an optimal transport based forecast model (credit Jonah Botvinick-Greenhouse - visiting graduate student).

High performance heterogeneous computing

Building a scalable and reproducible ecosystem for scientific machine learning research (Journal of Computational Science 62 (2022): 101750).