Lessons and Outlook for ML Parameterization of Sub Grid Atmospheric Physics From the Vantage of Emulating Cloud Superparameterization - Mike Pritchard

Explore lessons and future prospects for machine learning parameterization of sub-grid atmospheric physics from the perspective of emulating cloud superparameterization in this 42-minute conference talk. Delve into the challenges of global modeling, multiskill modeling, and GPU computing in climate science. Examine creative approaches to short simulations, course graining, and feature engineering. Analyze the tradeoffs, generalization strategies, and physical credibility of neural network models in atmospheric physics. Gain insights into hyperparameter tuning, missing information, and the importance of reporting failures in ML-based climate modeling. Conclude with a discussion on cognitive dissonance, excitement, and the future of machine learning in atmospheric science.

Introduction
Motivation
Turbulence
Global modeling
The challenge
Multiskill modeling
Global storm resolving models
A silly first attempt
Aerosol cloud indirect effects
Regionalization
GPU Computing
Creative Complexity
Short Simulations
Course Graining
Super Crude Architecture
Lessons emerging
Feature engineering
Separate processes
Microphysical rates
Example
Constraints
Tradeoffs
Generalization
Strategy
Preprint
Results
Physical Credibility
Hyperparameter Tuning
Missing Information
Neural Network Tuning
Summary
Cognitive dissonance
Excitement
Thank you
Maria
Reporting failures
Retraining neural networks
Sampling
Failures