1 of 10

Day 3: Tutorial on Forecasting El Niño / Southern Oscillation

​

Speakers: Ankur Mahesh

​

MC: Kelly Kochanski

​

​

___

CCAI Summer School

August 17, 2022

2 of 10

​

Ankur Mahesh

is a Ph.D. student at UC Berkeley in climate and atmospheric sciences. He uses methods from machine learning, fluid dynamics, and high-performance computing to constrain future projections of extreme weather. Ankur is also a core team member of Climate Change AI, where he is co-organizing the CCAI Summer School. Previously, Ankur worked on building climate resilience in agriculture with seasonal and long-term climate projections, and he was a member of the factory computer vision team at Tesla. He is a co-recipient of the Gordon Bell Prize for outstanding achievement in high-performance computing, and he has earned outstanding student presentation awards at the American Geophysical Union and American Meteorological Society annual conferences.

​

3 of 10

Day 3 Office Hours

  • Office hours today will be held right after the summer school from 18:10 to 20:00 UTC in the Room 2 Zoom: https://climatechange.ai/join/summer_school_room2
  • Please feel free to come with questions from this session or general questions about the summer school!

4 of 10

Setup

  • Please see Kureha’s post on Circle to open the link to the notebook for the tutorial today!
  • The title of the Circle post from this morning is “opening [DAY 3 - Climate Science]”
    • Press the link at the bottom of the post: “A reminder that the tutorial notebook has been pre-released here for those who would like to have a look at it before.”
    • then press the link for the tutorial on Day 3: “Forecasting the El Nino/ Southern Oscillation with Machine Learning”

5 of 10

Introduction to Tutorial on Forecasting El Niño/ Southern Oscillation

6 of 10

What is El Niño?

  • Cycle of warm and cold temperatures in the equatorial Pacific Ocean
  • Dominant pattern that influences seasonal temperature
  • Broad implications for climate-sensitive sectors, such as energy and agriculture
  • How is El Niño measured? Niño3.4 Index
    • Rolling 3-month average of sea surface temperatures in the equatorial Pacific

​

Source: National Oceanic and Atmospheric Administration

Equatorial Pacific Ocean with abnormally warm temperature: El Niño event

7 of 10

Learning to forecast El Niño

What is the current state of the art?

  • Most ENSO forecasts are issued by weather centers, who run physics-based models

​

Why use neural networks?

  • Potential for more accurate forecasts?
  • Lighter computational cost during inference
  • Challenge: limited historical observations to use as training data for a neural network
  • Solution: train on simulated climate data from Atmosphere-Ocean General Circulation Models (AOGCMs)

8 of 10

What questions will we explore during the tutorial?

  • Data: How does an increase in data affect the performance of machine learning?
  • Validation: how can we ensure that we validate the model rigorously?
  • Ensembling: What combination of models and training schemes creates the best forecasts?
  • Lead time: How far ahead can machine learning make skillful predictions?
  • Extendability: Can we use our neural network architecture to forecast temperatures on land?

Predictor Data: surface temperature

Target Data:

Source: NASA

Source: Columbia University

9 of 10

What might the forecasts look like? (4 month lead time)

SEAS5: seasonal forecasting model from the European Center for Medium-range Weather Forecasts

CNN+LSTM: a deep learning architecture designed to learn from spatial and time series data

CNN+LSTM

10 of 10

There’s still work to do on ENSO forecasting!

  • Why work on this problem?

​

​

​

​

  • Deep learning’s performance at extreme values of the Niño3.4 index still has room for improvement!

CNN+LSTM