The Machine Learning Packaging Paradigm: Optimising Climate Model Parameterisation Packaging

Climate models are essential tools for understanding and predicting the complex Earth system. Parameterisation is required to accurately simulate phenomena such as cloud formation, convection, and radiative transfer that occur on the sub-grid scale. Machine learning (ML) techniques present new opportunities to extract patterns directly from observational data or high-resolution simulations. However, they also present several challenges to ensuring accessible, reusable, reproducible and interoperable software and science. These include design and packaging of ML models, sharing of trained weights, and interoperability between ML frameworks and numerical models.

This poster showcases methods of tackling these challenges following the FAIR principles of software engineering. We present approaches to designing reusable machine learning pipelines with methods capable of accommodating distinct sets of input variables predicting the same target variables using adaptive data loading. Our work also addresses data heterogeneity by developing strategies to handle input variables differing in dimensionality.

We will discuss our pyTorch-Fortran coupling library FTorch facilitating interoperability between python ML and larger Fortran-based models. Another important feature of our work is saving and sharing the trained model weights using Hugging Face for efficient model distribution. We prioritise model reusability through a modular, extensible code design. Reproducibility is ensured via version control, and fixing random seeds. 

Our work showcases interdisciplinary collaboration between climate scientists, ML experts, and software engineers coupling ML parameterizations to the Community Earth System Model (CESM). Tackling these challenges helps improve the accuracy and predictive capabilities of climate models to inform climate change mitigation and adaptation strategies.