We develop machine learning methods and high-performance computing systems for scientific discovery. We study how machines learn physical dynamics and how these models can be made computationally scalable.
How can machines learn the structure and dynamics of physical systems?
We develop neural operators, generative models, transformers, and foundation models for scientific prediction. We study how these models represent spatiotemporal dynamics, generalize across physical regimes, and learn representations that extend beyond the conditions seen during training.
How can physical knowledge be incorporated into learning systems?
We develop learning-based models for fluid dynamics, heat transfer, multiphysics, and phase change systems. We investigate methods to encode conservation laws, geometry, and other physical constraints into learned models, and how to construct models whose predictions remain consistent with fundamental physical principles.
How can we make scientific AI scale?
We develop algorithms, GPU kernels, and efficient architectures for the computational bottlenecks underlying scientific AI. Our work spans attention, memory-efficient computation, and scalable training and inference, enabling models that would otherwise be computationally impractical.
Shakeel Hassan awarded the LANL-SoCal Hub Graduate Fellowship.
Aparna Chandramowlishwaran will give a Keynote at the 2027 Gordon Research Conference on Micro and Nanoscale Phase Change Phenomena.
Aparna Chandramowlishwaran will give a Keynote and serve as a Panelist at ASME InterPACK 2026.
Aparna Chandramowlishwaran promoted to Full Professor.
NUCLEUS-MoE was accepted to KDD 2026, a mixture-of-experts foundation model for pool boiling and liquid cooling. Publication →
HB-ARFM was accepted to ICML 2026, introducing history-bootstrapped flow matching for inverse boiling reconstruction. Publication →
