ATL Standard

A new machine learning technique developed at Georgia Tech promises to enhance weather forecasting and tsunami prediction. The method, known as Latent-EnSF, was created by Ph.D. student Phillip Si and Assistant Professor Peng Chen. It aims to improve data assimilation in predictive models.

Latent-EnSF has shown superior accuracy, convergence speed, and efficiency compared to existing methods in experiments related to medium-range weather forecasting and shallow water wave propagation. "We are currently involved in an NSF-funded project aimed at providing real-time information on extreme flooding events in Pinellas County, Florida," said Si.

The technique outperformed three other models during tests for shallow water wave propagation. These results suggest that Latent-EnSF can produce better predictions of coastal flood waves, tides, and tsunamis.

In addition to these tests, the model demonstrated scalability advantages over other methods when used for medium-range weather forecasting. Traditionally reliant on large supercomputers, such studies are becoming more feasible with smaller ML models like Latent-EnSF.

“Resolution, complexity, and data-diversity will continue to increase into the future,” said Chen. “To keep pace with this trend, we believe that ML models and ML-based data assimilation methods will become indispensable for studying large-scale complex systems.”

Data assimilation involves updating predictions with real-world data that is often sparse or incomplete. Latent-EnSF builds upon the Ensemble Filter Scores (EnSF) model developed by researchers at Florida State University and Oak Ridge National Laboratory.

The Georgia Tech team uses two variational autoencoders (VAEs) within Latent-EnSF to help integrate sparse data more accurately. This approach accelerates data assimilation processes and could provide faster predictions during crises.

Chen and Si presented their findings at the SIAM Conference on Computational Science and Engineering (CSE25), organized by the Society of Industrial and Applied Mathematics in Fort Worth, Texas. They plan further presentations at the International Conference on Learning Representations (ICLR 2025) in Singapore.

“We hope to bring attention to experts and domain scientists the exciting area of ML-based data assimilation by presenting our paper,” Chen said.