2025-09-11 All-Hands Presentation Meeting Notes
Toward Integrating E3SM with Machine Learning-enhanced Data Assimilation for Improved Earth System Predictability
Peyman Abbaszadeh
Department of Civil and Environmental Engineering, Portland State University
Abstract: Data assimilation plays a critical role in improving Earth system predictability by optimally integrating observations with models to reduce uncertainty and enhance forecasts. Advances in machine learning, high-performance computing, and novel assimilation techniques are now enabling new capabilities for prediction across weather, seasonal-to-decadal (S2D), and hydroclimate timescales. In this talk, I will highlight how these advances can be leveraged to improve predictions of hydrologic extremes such as floods and droughts, which remain pressing challenges for water resource management in a changing climate. The integration of remotely sensed observations into land surface and hydrologic models will be explored through state-of-the-art approaches, including ML-enhanced data assimilation. Recent developments in hydrologic data assimilation systems will also be discussed. I will present progress in hyper-resolution land surface and hydrologic data assimilation systems, emphasizing their role in strengthening predictive capabilities for water-related natural hazards and informing sustainable water resource management strategies. I will wrap up my talk with our collaboration with scientists from PNNL, where we are coupling ML-enhanced data assimilation with the Energy Exascale Earth System Model (E3SM) to improve the predictability of water and energy fluxes across the United States – a new development that has not been done before and is expected to significantly enhance the S2D predictability skill of E3SM.
Date Sep 11, 2025
Time
PT: 8:30 am
ET: 11:30 am
Call Info
web session: https://global.gotomeeting.com/join/570361173
call number: (571) 317-3122 Access Code: 570-361-173, If busy, use alternate number: (773) 945-1029
Joining from a video-conferencing room or system? Dial: 67.217.95.2##570361173 , Cisco devices: 570361173@67.217.95.2
Time | Title | Presenter | Presentation | Recording | Notes |
|---|---|---|---|---|---|
30 min | Toward Integrating E3SM with Machine Learning-enhanced Data Assimilation for Improved Earth System Predictability | Peyman Abbaszadeh | 
|
|
Recap of questions from Marcus van Lier-Walqui (Columbia University) and the responses from the speaker
Marcus: Nice talk! I was wondering what the relationship is between misspecification of state at the initial simulation time and fidelity of parameter estimation. If the state prior is really bad does it mess up paraemter estimation?
Peyman: Great question. The short answer is yes—misspecifying the initial state can negatively impact parameter estimation, especially early on. This is because the likelihood, which influences both the particle weights and the parameter updates, depends on the state and its corresponding observation. So, if the prior state is poor, it can lead to ineffective updates and a degraded posterior. However, sequential filtering is designed to handle this over time. Since we assimilate new observations at each time step, the system can gradually correct itself. In dual state-parameter estimation, the interaction between states and parameters helps reinforce this correction. If you suspect the initial state is poorly specified, it becomes even more important to ensure sufficient observational data is available. With enough observations, the filtering process can move both the state and parameters toward more accurate estimates. This is the power of Bayes' Law: at each step, we update the prior using the likelihood from the new observation to form a better posterior. Repeating this process over time allows the model to recover from a poor start, making the posterior increasingly reliable—even if the initial conditions were off.
Marcus: Relatedly, would you say the land system (while nonlinear) maybe does not have the chaotic characteristics of atmospheric processes?
Peyman: I'm not an atmospheric scientist, but from what I understand, land surface processes are definitely nonlinear, though probably not as chaotic as atmospheric ones. The atmosphere is really sensitive to initial conditions—which is a key feature of chaotic systems—so small differences can lead to big changes really quickly. Land systems, on the other hand, tend to evolve more slowly and are generally more constrained, so they don’t show that same level of chaotic behavior. That said, land processes are still complex, with lots of interactions, feedback loops, and thresholds (like soil moisture or vegetation changes), so modeling them is still challenging—just in a different way.
Happy to discuss this more.