Machine learning workflows in climate modelling: design patterns and insights from case studies.

Area of Science:

• Climate Science
• Data Science
• Computational Science

Background:

• Machine learning (ML) is increasingly vital in climate modeling for tasks like emulation, parameter inference, and forecasting.
• Challenges include physical consistency, multi-scale coupling, data sparsity, generalization, and workflow integration.

Purpose of the Study:

• To analyze case studies of applied ML in climate modeling, focusing on workflow design patterns.
• To synthesize diverse ML approaches, including surrogate modeling, parameterization, probabilistic programming, simulation-based inference, and physics-informed transfer learning.
• To demonstrate a framework for rigorous and reproducible scientific ML in climate science.

Main Methods:

• Analysis of case studies in ML-enabled climate modeling.
• Synthesis of workflow design patterns across various ML applications.
• Demonstration of a framework for transparent model development, evaluation, adaptation, and reproducibility.

Main Results:

• Identification of key workflow design patterns for ML in climate modeling.
• Demonstration of how ML workflows can be grounded in physical knowledge, simulation data, and observations.
• A framework is presented to ensure scientific rigor and facilitate interdisciplinary collaboration.

Conclusions:

• ML offers powerful tools for advancing climate modeling, but careful workflow design is crucial.
• The proposed framework enhances transparency, reproducibility, and collaboration in ML-driven climate science.
• Lowering barriers for data science and climate modeling integration is essential for future progress.