Overview: We examine climate emulators and out-of-sample predictions.

Models Used

• 1. Climate in a Bottle (cBottle)

  A generative diffusion model developed by NVIDIA. It uses a coarse-generation model (~100 km resolution) trained on ERA5 (1980-2017). Uniquely, it is non-autoregressive, generating hourly outputs directly from SST inputs. We evaluate 10-year control and +2K simulations.

• 2. ACE2-ERA5

  An autoregressive ML model using Spherical Fourier Neural Operators. Trained on ERA5, it operates at 1° spatial and 6-hourly temporal resolution. It combines a convolutional encoder with a Transformer-based predictor. We perform 10-year simulations initialized from 2001 conditions.

• 3. NeuralGCM

  A hybrid model combining a differentiable dynamical core with neural network parameterizations. Unlike pure ML emulators, it integrates physical dynamics with learned physics. It produces stable decadal simulations (140–280 km resolution) and includes precipitation outputs.