PyEarthTools Training Day

Machine Learning for Climate and Weather
1

This post will hold the schedule for the ACCESS-NRI 2025 training day using PyEarthTools to introduce people to machine learning for climate and weather. It will be updated progressively as we get closer to the event.

Untitled
Untitled164×164 1.83 KB

QR code directing to the URL for this Hive Forum post ( PyEarthTools Training Day )

If you want to get started ahead of time, the tutorial will be based on these linked notebooks: https://github.com/ACCESS-Community-Hub/PyEarthTools/tree/develop/notebooks/tutorial/ENSO_Tutorial

For convenience, a copy of these tutorials is on disk at /scratch/nf33/pyearthtools_notebooks which should be accessible to logged-in users. Make a copy of these notebooks in a directory of your choosing (such as your home directory, or your user-specific directory under /scratch/nf33) and work from there.

You will need to specify the following information when launching a Jupyter session from https://are.nci.org.au/ .

Walltime (hours): 4

Queue: normalbw

Compute Size: large

Storage: gdata/dk92+gdata/rt52

Extra arguments:

Module directories: /g/data/dk92/apps/Modules/modulefiles

Modules: pet/2025.08

Jobfs size: 100GB

1.45 - 3.15: Connecting to data with PET; problem overview, XGBoost example

3.30 - 5pm: Neural networks with PET

Session 1 - Setup, data, xgboost

1.45 – 2.15pm: Get everyone online, connected and set up, join projects. Run through the basic plotting notebook as proof of success

2.15 – 2.35: Introduction to ENSO, why is it important, what prior ML work exists

2.35 – 2.50: Introduction to data accessors in PET

2.50 – 3.15: XGBoost model of ENSO with PET:

  • Data overview
  • Data preparation
  • Test/train/validate
  • Connection to XGBoost
  • Introduction to PyEarthTools pipeline features
  • If time, how to add new features using additional data accessors
    • Try changing the bounding box
    • Try adding an additional data source (e.g. satellite or another model variable)
    • Other next steps from the notebooks as time allows

3.15 – 3.30: Afternoon tea and tech support

Session 2 – Neural Networks and Model Evaluation

3.30 – 4pm: Introduction to neural networks with PET

  • Separating the input pipeline from the target pipeline
  • Preserving geospatial structures
  • Sequence-to-sequence pipelines
  • Normalising (z-scaling) inside a pipeline across dimensions
  • Presenting data to PyTorch

4pm-4.30pm: Model training and evaluation

  • Training larger models on GPUs
  • Exploring loss functions and training strategies

4pm – 4.30pm: Neural network architectures for ENSO

  • Introduction to CNNs and ResNet
3 Likes