xaig

Light, framework-agnostic tooling for E3SM AI campaigns: what an emulator holds inside, sparse autoencoders trained on it, figures, and a local app over them.

Research tool. xaig is early. What is described here works; expect it to change.

Install

The base install pulls only Click. Anything heavier sits behind an extra named after the subpackage that needs it:

$ uv pip install 'xaig[daig]'     # or: pip install 'xaig[daig]'

A missing extra says so, with the command that fits how xaig was installed.

Try it, with no model and no data

A toy emulator writes a latent archive with nothing but numpy; everything else reads it like any other:

$ xaig daig latent toy scratch/toy/control
$ xaig daig latent info scratch/toy/control --mask-variable sst
$ xaig daig latent region scratch/toy/control --lat 10 --lon -114 --time 2 --centred --pcs 2
$ xaig daig latent fields scratch/toy/control --field precipitation --top 3
$ xaig taig sae scratch/toy/control --features 64 --k 4 --out scratch/toy/sae.npz   # needs xaig[taig]
$ xaig waig --latents scratch/toy                                                  # needs xaig[waig]

The CLI is a thin client of the Python API; anything it can do, a notebook can:

from xaig.daig.latent import Region, analyse_region, open_source

source = open_source("scratch/toy/control", mask_variable="sst")
result = analyse_region(
    source, time=2, layer=3, region=Region(lat=10, lon=-114, radius_km=1500), centred=True
)
result.ranking.channels  # the channels that respond most strongly there

More

• Guides: https://e3sm-project.github.io/aigroup/package/
• Source and issues: https://github.com/E3SM-Project/aigroup

BSD-3-Clause. xaig.daig.latent, xaig.faig and xaig.waig grew out of the latent space visualiser for weather models (Tempest, Beylich & Craig 2026, arXiv:2604.20467, doi:10.1007/978-3-032-29915-4_10); see NOTICE. The sparse autoencoders follow MacMillan & Ouellette (2025, arXiv:2512.24440); the B-spline autoencoder of Cheon (2026, arXiv:2605.17493) is what xaig.taig is heading for and does not implement yet.