Research Questions¶
This is inspired by the talk by Steve Penny - Recording | Slides
Reanalysis vs Simulations vs Observations¶
Observations and Reanalysis are inherently imperfect data sources with often uncharacterized uncertainties.
- Q1: Are reanalysis datasets an adequate source of training data for ML?
- Q2: Are pure simulation datasets more effective data for ML?
- Q3: How will biases & systematic errors be handled?
- Q4: Can we learn directly from observations plus basic physics constraints?
Reanalysis-Based
Simulation-Based
PINNS in the Wild
Hybrid Models¶
As numerical Forecasts are modernized (e.g. written in new languages that support differentiation, and designed to take advantage of GPUs), can AI/ML solutions maintain a competitive edge (in terms of computational cost) over conventional modeling.
[Belochitski & Krasnopolsky (2021)Dresdner et al. (2022)Frerix et al. (2021)Kochkov et al. (2021)]
Model Error Estimation¶
How much State Dependent (Conventional) Model error can we learn from comparison with observations? How do we separate system observation errors from systematic model forecast errors?
[Bonavita & Laloyaux (2022)Laloyaux et al. (2022)Pathak et al. (2018)Arcomano et al. (2022)]
Subgrid Parameterization¶
This is an instance of
Better Metrics¶
Operational Center¶
Observation Datasets¶
AlongTack
In-Situ
- Currents - European Union-Copernicus Marine Service, 2018
- Stuff - European Union-Copernicus Marine Service, 2015
- ARGO - Wong et al. (2020)Argo (2023)
Extrapolation Datasets¶
Forecast
- OceanPhysics - European Union-Copernicus Marine Service, 2016 | Ensemble - European Union-Copernicus Marine Service, 2019
- Ocean Biogeochemistry - European Union-Copernicus Marine Service, 2019
- Ocean Waves - European Union-Copernicus Marine Service, 2018
HindCast
Reanalysis Datasets¶
- European Union-Copernicus Marine Service, 2018
- CDS - Copernicus Climate Change Service, 2021 | SST - Copernicus Climate Change Service, 2019
- NCEP - 10.1175/1520-0493(1998)126<1013:AICMFE>2.0.CO2
- NCI - Chamberlain et al. (2021)Chamberlain et al. (2021)
- List of Data
Applications¶
Plants n Things¶
- Hanna Meyer - Slides | Videos
- Nature Conservation - Tuia et al., 2022
- Agriculture
Water n Things¶
- Maritime Risk Predictions - Knapp & Velden (2023)Rawson & Brito (2022)
- IoT - example
- Monitoring Ocean Health - Longo et al. (2017)Franke et al. (2020)
- Underwater Exploration and Ocean Cleanup
- Sustainable Fishing Practices
- Marine Protected Areas
- Ocean Energy Solutions
Data Assimilation¶
Algorithms¶
Back-and-Forth Nudging¶
- Agarwal, A., Meijer, V. R., Eastham, S. D., Speth, R. L., & Barrett, S. R. H. (2022). Reanalysis-driven simulations may overestimate persistent contrail formation by 100%–250%. Environmental Research Letters, 17(1), 014045. 10.1088/1748-9326/ac38d9
- Belochitski, A., & Krasnopolsky, V. (2021). Robustness of neural network emulations of radiative transfer parameterizations in a state-of-the-art general circulation model. Geoscientific Model Development, 14(12), 7425–7437. 10.5194/gmd-14-7425-2021
- Dresdner, G., Kochkov, D., Norgaard, P., Zepeda-Núñez, L., Smith, J. A., Brenner, M. P., & Hoyer, S. (2022). Learning to correct spectral methods for simulating turbulent flows. arXiv. 10.48550/ARXIV.2207.00556
- Frerix, T., Kochkov, D., Smith, J. A., Cremers, D., Brenner, M. P., & Hoyer, S. (2021). Variational Data Assimilation with a Learned Inverse Observation Operator. arXiv. 10.48550/ARXIV.2102.11192
- Kochkov, D., Smith, J. A., Alieva, A., Wang, Q., Brenner, M. P., & Hoyer, S. (2021). Machine learning–accelerated computational fluid dynamics. Proceedings of the National Academy of Sciences, 118(21). 10.1073/pnas.2101784118