Field-space autoencoder for scalable climate emulators

Abstract High-Resolution Earth system models are essential for capturing local climate change. However, these models are computationally expensive and produce petabyte-scale outputs, limiting their utility for applications such as probabilistic risk assessment. Here, we present the Field-Space Autoencoder, a scalable climate emulation framework based on a spherical compression model that overcomes these challenges. By combining a multi-scale encoding of the input with field-space processing layers, the model efficiently processes multiple spatial scales simultaneously. This approach preserves physical structures significantly better than convolutional baselines. By producing a structured compressed field, it serves as a good baseline for downstream generative emulation. In addition, the model can perform zero-shot super-resolution that maps low-resolution large ensembles and scarce high-resolution data into a shared representation. By training a generative diffusion model on these compressed fields, we demonstrate that our framework can simultaneously learn internal variability from abundant low-resolution data and fine-scale physics from scarce high-resolution data. Our work bridges the gap between the high volume of low-resolution ensemble statistics and the scarcity of high-resolution physical detail.