license: mit
tags:
- climate
- weather-forecasting
- swin-transformer
- flow-matching
ArchesClimate-SSP
Deep-learning climate emulator (Swin-Transformer-based, autoregressive) trained to reproduce IPSL-CM6A-LR and CanESM5 monthly climate states conditioned on external forcings (GHGs, aerosols, ozone, solar irradiance). Generates unseen SSP scenario trajectories cheaply, without rerunning the full Earth System Model.
Code: https://github.com/gclyne/ArchesClimate (see INFERENCE.md there for
the full inference recipe).
What's in this repo
Only what's needed to instantiate and load the network — weights + the Hydra config each checkpoint was trained with:
<run_name>/
step-step=NNNNNN.ckpt # model weights (raw + EMA)
config.yaml # exact training-time Hydra config
Required companion data
This repo does not include initial conditions, forcing (boundary condition) trajectories, static fields, or normalization stats — those live in the paired dataset repo: gclyne/ArchesClimateDataset.
See INFERENCE.md in the code repo for exactly which files from that
dataset repo are needed for a given checkpoint, and how to run a rollout.
Runs currently published here
| Run | Steps | Notes |
|---|---|---|
pf2_emafix_step40000 |
40,000 | Pushforward-length-2 energy-score deterministic model, EMA weights |