Global AI Accelerator - GAIA

Welcome to GAIA Team Project Page, a DARPA AIE ACTM project.

Code repository and documentation for:

training NN surrogates: https://github.com/stresearch/gaia
deploying NN surrogates in GCM (still private, will be made public soon): https://github.com/stresearch/gaia-deploy

Project Description

STR
- Jim Gimlett
- Gregory Bushey
- Kirill Trapeznikov
- Eric Jones
University of New South Wales, Sydney
- Steven Sherwood
- Abhnil Prasad
- David Fuchs

Datasets - description of datasets used in surrogate training
Surrogate Details and Hyperparameter Sweeps - neural network architecture and hyperparameter comparison
Comparison of Surrogates Trained on Different Datasets - compare surrogates to CAM4, SPCAM simulations and compare CAM4 and SPCAM4 trained surrogates to each other
Baseline Input Ablation - ablation of input variables and their effect on model performance
Surrogate with Memory Inputs - improving surrogate performance by adding memory i.e. outputs from previous timestep
Integration of AI Surrogate into GCM - deploying GAIA ML model in the GCM
MJO Analysis - analysis of MJO in the datasets
Surrogate Bottleneck Analysis - analysis of bottleneck dimension of the surrogate model
Linearization - analysis linearization of surrogate models to understand sensitivity to inputs
Surrogate Architectures - comparison of different neural network architectures and configurations
Fine Tuning - comparison of fine-tuning an CAM4 trained model on SPCAM vs a model traiend from scratch
Integration Journal - details on integration including training setup, diagnostic plots, etc.

Animation of CAM4 Simulation and Neural Network Surrogate Model

Concept & Approach

Develop new hybrid AI tools and methods to accelerate ultra-high-resolution Global Climate Models (GCMs) models, substituting AI local surrogates to improve cloud-resolving physics
Test climate “tipping point” hypotheses, with focus on the Madden Julian Oscillation (MJO)

What’s hard?

GCMs are computationally expensive and lack the resolution needed to adequately model local convection and thus clouds
This greatly increases GCM forecast errors and impedes propagation of large-scale wave phenomena such as the MJO

What will GAIA accomplish?

Enable a GCM to accurately model local convection and predict self-organizing atmospheric wave phenomena
Exploit this GCM to explore possible future regimes and identify early warning signatures for MJO-related tipping points.

Progress

Developed and validated high-skill AI surrogates for multiple GCM local models, including CAM4, SPCAM with Morrison micro-physics, SPCAM with CLUBB cloud parameterization
Currently studying impact of added memory and ablating inputs
Next Steps:
- Determine input regimes where CAM/SPCAM underperform
- Analyze GCM sensitivity using lower order dynamical model
- Apply results to selectively enhance local surrogate using ultra-high-resolution local eddy simulation for training.
- Integrate surrogate into GCM; use GCM to test MJO hypotheses