Multi-stage ensemble-learning-based model fusion for surface ozone simulations: A focus on CMIP6 models.

Area of Science:

• Atmospheric Chemistry and Climate Science
• Computational Modeling
• Machine Learning Applications

Background:

• Accurate simulation of global surface ozone distribution and variability is crucial for chemistry-climate modeling.
• Significant discrepancies exist in current model outcomes due to uncertainties in tropospheric ozone budget factors.
• There is a need for advanced methods to overcome structural biases and improve surface ozone prediction accuracy.

Purpose of the Study:

• To develop and evaluate machine learning approaches for fusing multiple climate model simulations to enhance surface ozone prediction.
• To address cross-model discrepancies and improve the fidelity of chemistry-climate model outputs.

Main Methods:

• Utilized the Coupled Model Intercomparison Project Phase 6 (CMIP6) ensemble of 57 simulations.
• Implemented a conventional ensemble learning approach.
• Developed an innovative 2-stage enhanced space-time Bayesian neural network.

Main Results:

• Both machine learning approaches achieved outstanding performance (R² > 0.95, RMSE < 2.12 ppbv).
• The conventional ensemble learning approach was computationally cheaper with higher overall performance but lacked interpretability and overestimated oceanic ozone.
• The Bayesian approach demonstrated superior spatial generalization and interpretability but incurred higher computational costs.

Conclusions:

• Multi-stage machine learning frameworks can significantly improve the accuracy of chemistry-climate model outputs.
• These enhanced models provide a more reliable basis for future climate impact studies.
• The choice between ensemble learning and Bayesian networks depends on computational resources and interpretability requirements.