Machine Learning Encoder Improves Weather Forecasting and Tsunami Prediction

Successful test results of a new machine learning (ML) technique developed at Georgia Tech could help communities prepare for extreme weather and coastal flooding. The approach could also be applied to other models that predict how natural systems impact society. 

Ph.D. student Phillip Si and Assistant Professor Peng Chen developed Latent-EnSF, a technique that improves how ML models assimilate data to make predictions.

In experiments predicting medium-range weather forecasting and shallow water wave propagation, Latent-EnSF demonstrated higher accuracy, faster convergence, and greater efficiency than existing methods for sparse data assimilation.

“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, who studies computational science and engineering (CSE). 

“We're actively working on integrating Latent-EnSF into the system, which will facilitate accurate and synchronized modeling of natural disasters. This initiative aims to enhance community preparedness and safety measures in response to flooding risks.” 

Latent-EnSF outperformed three comparable models in assimilation speed, accuracy, and efficiency in shallow water wave propagation experiments. These tests show models can make better and faster predictions of coastal flood waves, tides, and tsunamis. 

In experiments on medium-range weather forecasting, Latent-EnSF surpassed the same three control models in accuracy, convergence, and time. Additionally, this test demonstrated Latent-EnSF's scalability compared to other methods.

These promising results support using ML models to simulate climate, weather, and other complex systems.

Traditionally, such studies require employment of large, energy-intensive supercomputers. However, advances like Latent-EnSF are making smaller, more efficient ML models feasible for these purposes.

The Georgia Tech team mentioned this comparison in its paper. It takes hours for the European Center for Medium-Range Weather Forecasts computer to run its simulations. Conversely, the ML model FourCastNet calculated the same forecast in seconds.

“Resolution, complexity, and data-diversity will continue to increase into the future,” said Chen, an assistant professor in the School of CSE. 

“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 is the process by which models continuously ingest new, real-world data to update predictions. This data is often sparse, meaning it is limited, incomplete, or unevenly distributed over time. 

Latent-EnSF builds on the Ensemble Filter Scores (EnSF) model developed by Florida State University and Oak Ridge National Laboratory researchers. 

EnSF’s strength is that it assimilates data with many features and unpredictable relationships between data points. However, integrating sparse data leads to lost information and knowledge gaps in the model. Also, such large models may stop learning entirely from small amounts of sparse data.

The Georgia Tech researchers employ two variational autoencoders (VAEs) in Latent-EnSF to help ML models integrate and use real-world data. The VAEs encode sparse data and predictive models together in the same space to assimilate data more accurately and efficiently.

Integrating models with new methods, like Latent-EnSF, accelerates data assimilation. Producing accurate predictions more quickly during real-world crises could save lives and property for communities.

[Related: University of South Florida Researchers Track Flooding in Coastal Communities During Hurricanes Helene and Milton]

To share Latent-EnSF to the broader research community, Chen and Si presented their paper at the SIAM Conference on Computational Science and Engineering (CSE25). The Society of Industrial and Applied Mathematics (SIAM) organized CSE25, held March 3-7 in Fort Worth, Texas.

Chen was one of ten School of CSE faculty members who presented research at CSE25, representing one-third of the School’s faculty body. Latent-EnSF was one of 15 papers by School of CSE authors and one of 23 Georgia Tech papers presented at the conference.

The pair will also present Latent-EnSF at the upcoming International Conference on Learning Representations (ICLR 2025). Occurring April 24-28 in Singapore, ICLR is one of the world’s most prestigious conferences dedicated to artificial intelligence research.

“We hope to bring attention to experts and domain scientists the exciting area of ML-based data assimilation by presenting our paper,” Chen said. “Our work offers a new solution to address some of the key shortcomings in the area for broader applications.”