ACE2S-SHiELD-plus-supplemental-checkpoints

<img src="ACE-logo.png" alt="Logo for the ACE Project" style="width: auto; height: 50px;">

ACE2S-SHiELD+ supplemental checkpoints

This repository contains a collection of checkpoints associated with supplemental models trained and referenced in the ACE2S-SHiELD+ manuscript. These checkpoints may be useful for certain research purposes, but for almost all cases we recommend using the ACE2S-SHiELD+ checkpoint from this repository, which contains the model featured in the manuscript, which will likely produce the best results.

The models here were trained for ablation experiments and/or with alternative random seeds. The ablation experiments aimed to separate the impacts of including random CO₂ data in training and imposing energy conservation. Accordingly models with four different configurations are included in separate subdirectories, which are described in the table below. For the technical details regarding what these different configuration options mean, please see discussion in the manuscript.

Two models with different random seeds were trained for each configuration. Within each subdirectory they are labeled rs0_ckpt.tar and rs1_ckpt.tar. For each model, the checkpoint was chosen from the epoch with best inline inference skill. In the manuscript, unless results from both seeds are shown, for visual clarity results from the model with best inline inference skill for the particular configuration are used, which is denoted in the table above. In all cases, inline inference consisted of an ensemble of simulations in 1x, 2x, and 4xCO₂ equilibrium climates, and the error was computed based on the approach described in Watt-Meyer et al. (2025).

*This is the checkpoint featured in the main ACE2S-SHiELD-plus repository. Therefore it is not included in this repository.

License

These models are licensed under Apache 2.0. They are intended for research and educational use in accordance with Ai2’s Responsible Use Guidelines.