Home

Our group is moving to the University of Chicago. We have a number of PhD/postdoc positions for Spring/Fall 2024 to work at the intersection of extreme weather, scientific machine learning, climate change, and computational physics. See Available Positions for more details.

We study extreme weather, climate change, and geophysical turbulence through the lens of multi-scale nonlinear dynamics. We aim to integrate tools and concepts from fluid dynamics, applied math, scientific ML, and computational physics to gain a deeper theoretical understanding of these phenomena and develop better computational tools to predict them. Examples of problems of interest are blocking events, heat waves and hurricanes, subgrid-scale modeling of geophysical turbulence and gravity waves, annular modes, and explainable physics-informed ML. Our research has been supported by NSF, ONR, NASA, Schmidt Futures (VESRI), National Academy of Sciences, C3.ai Digital Transformation Institute, Accenture, Microsoft AI, Mitsubishi Electric Research Lab, Rice University Creative Ventures, and Rice Houston Engagement and Recovery Effort.

Recent News:

  • October 2023: Check out the recording of the talk titled “Integrating physics, data and scientific machine learning to predict climate variability and extremes” I gave as a part of the APS-GPC seminar series.
  • May 2023: Check out the recording of the talk titled “Opening the neural networks’ black box for climate science” I gave at AI for Good seminar series.
  • February 2023: Paper “Explaining the physics of transfer learning in data-driven turbulence modelingled by then undergrad student Adam Subel is published in Proceedings of the National Academy of Sciences Nexus. In this paper, we introduced a framework based on integrating the Fourier analyses of deep convolution neural networks (CNNs) and data to fully explain the physics learned by the CNNs when applied to multi-scale nonlinear dynamical systems such as turbulence and climate. We demonstrated how millions of learned parameters reduce to a few classes of well-known spectral filters. We also showed how transfer learning enables out-of-distribution generalization to vastly different flows e.g., with 100x higher Reynolds numbers. See the Rice press release.
  • November 2022: Check out the recording of the talk titled “Learning Data-driven Subgrid-Scale Models for Geophysical Turbulence: Interpretation, Stability and Extrapolation” I gave at Columbia University’s LEAP Lectures on Climate Data Science.
  • October 2022: Paper “Learning physics-constrained subgrid-scale closures in the small-data regime for stable and accurate LES led by postdoc Yifei Guan is published at Physica D: Nonlinear Phenomena. In this paper, we showed how incorporating physics (e.g., symmetries, conservation laws) into the learning process of data-driven subgrid-scale parameterizations lead to stable and accurate large-eddy simulation using 40x fewer training samples.
  • March 2022: Paper “Towards physics-inspired data-driven weather forecasting: integrating data assimilation with a deep spatial-transformer-based U-NET in a case study with ERA5led by PhD student Ashesh Chattopadhyay is published at Geoscientific Model Development. In this paper, we integrated an equivariant U-Net with an ensemble Kalman filter to do data-driven weather forecasting with efficient data assimilation. We also devised a multi-time-resolution scheme to increase the accuracy across time scales.
  • March 2022: Paper Stable a posteriori LES of 2D turbulence using convolutional neural networks: Backscattering analysis and generalization to higher Re via transfer learningled by postdoc Yifei Guan is published at the Journal of Computational Physics. In this paper, we used CNNs for data-driven subgrid-scale (SGS) modeling, and particularly, showed that 1) backscattering is harder to learn (compared to diffusion), which can lead to instabilities in the small-data regime, 2) transfer learning enables extrapolation in Re. Results can help the ongoing efforts in developing data-driven SGS models for turbulent flows in various disciplines.

Clicky