uturuncoglu/

aurora

This repository has a version of Microsoft's Aurora Model that has been fine-tuned to make it easier to couple with traditional physics-based earth system model components, such as the Regional Ocean Modeling System (ROMS) and the Modular Ocean Model (MOM6). The model is fine-tuned using the fifth-generation ECMWF reanalysis for the global climate and weather dataset (ERA5) provided by ECMWF. Please visit my GitHub repository to access the code used to train the model, utilities to process and prepare the data for training, and learn how to use the checkpoints to make predictions.

Aurora is a 1.3 billion parameter foundation model for high-resolution forecasting of weather and atmospheric processes. Aurora is a flexible 3D Swin Transformer with 3D Perceiver-based encoders and decoders. At pretraining time, Aurora is optimized to minimize a loss on multiple heterogeneous datasets with different resolutions, variables, and pressure levels. The model is then fine-tuned in two stages: (1) short-lead time fine-tuning of the pretrained weights and (2) long-lead time (rollout) fine-tuning using Low Rank Adaptation (LoRA). The fine-tuned models are then deployed to tackle a diverse collection of operational forecasting scenarios at different resolutions.

More information about the model can be found in following links:

The Aurora model is fine-tuned to make prediction for following variables:

Note: The weights indicated with * are used with the checkpoint file (3) given in the next section.

The code used to train and validate results can be found in the following GitHub repository.