Climate Modeling with Generative AI

Climate Prediction: Trajectory or Distribution?

In machine learning weather assimilation, scientists use observations to adjust their models' trajectory and forecasts, such as daily temperatures. This traditional method focuses on short-term accuracy of the model. However, for climate models, it is also important to understand long-term dynamics and statistics in the state as we cannot replicate exactly a particular day in the future. This work introduces a new approach to improve data assimilation targeting statistics, by looking at the overall range of possible state values over time. This work also introduce a framework that allows us to infer model parameters while considering uncertainty, improving the reliability of our projections. Finally, it shows how this approach could help estimate extreme events such as extremely hot and cold days and analyze the distribution of the joint quantile values.

Accepted Geophysical Research Letters, 2025

Sparse Remote Sensing Observations and Probabilistic Climate Projection

This work aims to enhance climate projections by integrating transport processes into ESM. As a first step, I utilize indirect meteorological data such as gas flux, water vapor, and precipitation, processed through a machine learning-based data assimilation framework that I developed. This framework, designed for a multimodal setting, leverages deep learning to overcome the limitations of traditional data assimilation techniques that rely on linear and Gaussian assumptions. Using in-situ data from the CPC global rain gauges and ex-situ data from NOAA satellites, my approach refines global temperature profiles effectively, even under conditions of sparse and noisy data. Moreover, I have developed a probabilistic data assimilation method using probability density distributions for enhanced parameter inference and uncertainty quantification in a Bayesian framework. This fresh methodology is pivotal for projecting long-term climate scenarios, allowing for the estimation of extreme values and detailed analysis of significant climate events like flooding and heatwaves.

Published in CVPR- Earth Vision, 2024

This work pioneered a physics informed machine learning framework for modeling suspended sediment in vegetated flows, achieving a data-driven closure of the sediment dispersion flux. This expands traditional hydraulic models and establishes a direct link between microscale dispersive flux models and observed sediment concentration in channel flumes. Expanding beyond sediment flux distribution, my work includes the development of machine learning-assisted turbulent models to delineate the pier scouring process, which has implications for river channel protection engineering. This framework successfully links turbulent physics with prognostic equations for regional-scale scour predictions. By integrating the TKE and shear stress equations, it implicitly relates pier scouring depth to the largest turbulence eddy scale. However, given the complexity and analytical challenges involved, a traditional analytical solution remains elusive. Instead, my use of machine learning-based symbolic regression has demonstrated potential to unlock new insights, illustrating how turbulent eddies significantly influence the bulk-scale pier scouring process, thus welding AI with physical sciences to offer new understanding of these environmental interactions.

Published in Geophysical Research Letters, 2024

Bridge Microscale and Environment via AI

Page updated

Report abuse