MR4423 Data-Driven Weather Prediction

Introducing to machine learning weather prediction, including neural network architectures, ensemble forecasting methods, hybrid ML-numerical weather prediction systems, and global and mesoscale ML models. Students evaluate model performance, limits of predictability, uncertainty, and calibration. Through case studies, real-time forecasting, article discussions, and hands-on model inference using cloud and local GPU and/or TPU resources, students gain practical experience applying and critically assessing emerging machine learning weather prediction capabilities relevant to operational meteorological and oceanographic Fleet decision support.

Prerequisite

MR4323 or OC4323 or permission of instructor

Lecture Hours

3

Lab Hours

0

Course Learning Outcomes

1. Explain the differences between different neural network architectures used in

weather and climate prediction.

2. Assess and describe limits of predictability of data-driven ML weather prediction

models and compare these to those from numerical weather prediction (NWP)

models.

3. Quantify ensemble dispersiveness of data-driven ML ensemble weather prediction

systems for cyclone tracks and selected surface and upper-air variables, such as

wind, temperature, geopotential heights, etc.

4. Articulate the strengths and weaknesses of traditional NWP and ML weather

prediction at forecasting extreme weather events.

5. Describe methods used to generate ML ensembles for weather prediction, including

strengths and weaknesses of each method.

6. Critically evaluate articles in the research area of machine learning weather

prediction.

7. Run ML weather model inference in Google Cloud and/or on local GPUs using

Earth2Studio.