Machine Learning Workflows in Climate Modeling: Design Patterns and Insights from Case Studies

TL;DR

This work analyzes how machine learning can be integrated into climate modeling through modular, science-guided workflows that couple physics, data, and ML. It delineates four guiding directions—Physics-First, Data-First, ML-First, and Human-in-the-Loop—and demonstrates eight case studies ranging from neural-operator surrogates to simulation-based inference and transfer learning for sparse observations. The authors emphasize reproducibility, interpretability, online stability, and uncertainty quantification as core design goals, and propose structured design, development, deployment, and evaluation phases to achieve scientifically robust ML workflows. The paper highlights recurring patterns such as modularity, physics-informed constraints, and the crucial distinction between offline skill and online deployment performance, while outlining challenges in standardization, scalability, and deeper integration with scientific reasoning.

Abstract

Machine learning has been increasingly applied in climate modeling on system emulation acceleration, data-driven parameter inference, forecasting, and knowledge discovery, addressing challenges such as physical consistency, multi-scale coupling, data sparsity, robust generalization, and integration with scientific workflows. This paper analyzes a series of case studies from applied machine learning research in climate modeling, with a focus on design choices and workflow structure. Rather than reviewing technical details, we aim to synthesize workflow design patterns across diverse projects in ML-enabled climate modeling: from surrogate modeling, ML parameterization, probabilistic programming, to simulation-based inference, and physics-informed transfer learning. We unpack how these workflows are grounded in physical knowledge, informed by simulation data, and designed to integrate observations. We aim to offer a framework for ensuring rigor in scientific machine learning through more transparent model development, critical evaluation, informed adaptation, and reproducibility, and to contribute to lowering the barrier for interdisciplinary collaboration at the interface of data science and climate modeling.

Figures (12)

• Figure 1: Interacting roles of physics, data, and machine learning in climate modeling. Climate models, especially Earth System Models (ESMs), interact with two foundational sources of information: physics and data (simulated or observed). Machine learning is increasingly used to improve the models' reliability and utility. Physics informs model structure, while models in turn help discover new scientific insights. Data constrain model behavior, and models help emulate climate data, interpret observations, and infer unmeasured physical states. Machine learning contributes by improving model components and enabling new approaches to represent the climate system’s variability and dynamics.
• Figure 2: Machine learning workflow for climate modeling. Applying machine learning to Earth system modeling and climate data analysis spans three phases: design, development, and deployment. At each stage, domain knowledge and physical consistency guide decisions about model structure, objective functions, and evaluation. The process often iterates between design and development. Model integration, evaluation, and interpretation then lead to operational and scientific insights, supporting adaptive problem refinement and co-design in the workflows. Case studies that provide in-depth discussions of design choices are labeled in red (case study ID). This diagram is not meant to be a novel contribution, but rather to offer a familiar machine learning framework used to organize and contextualize our discussion.
• Figure 3: Data generating process for the Climsim dataset (annotated Figure 1 from benedict_structure_2009). The Climsim data set aims to support the development of ML emulators for Cloud Resolving Model (CRM) that could be readily integrated into operational climate simulators such as the General Circulation Model (GCM). A.) Begin with macro-state of the entire grid cell, $q_{\,G}^n$ for tracked variables. B.) $q_G$ updates the internal convective state of the cloud model and supplies boundary conditions. C.) The cloud model steps forward in time, matching the 20 minute interval of the simulation. D.) A coarsened average of cloud internals yields output $\langle q_C^M \rangle_\alpha$ with resolution matching the GCM. Atmospheric states before and after cloud resolution are saved as input-output vectors for the dataset. E.) Other processes and parameterizations are applied, cumulatively creating the next GCM macro state $q_{\,G}^{n+1}$.
• Figure 4: Workflow for machine learning-based parameterization of subgrid processes.
• Figure S1: Workflow for training and deploying ML-based surrogate models.