Abstract
Accurately simulating the geographical distribution and temporal variability of global surface ozone has long been one of the principal components of chemistry-climate modelling. However, the simulation outcomes have been reported to vary significantly as a result of the complex mixture of uncertain factors that control the tropospheric ozone budget. Settling the cross-model discrepancies to achieve higher accuracy predictions of surface ozone is thus a task of priority, and methods that overcome structural biases in models going beyond naïve averaging of model simulations are urgently required. Building on the Coupled Model Intercomparison Project Phase 6 (CMIP6), we have transplanted a conventional ensemble learning approach, and also constructed an innovative 2-stage enhanced space-time Bayesian neural network to fuse an ensemble of 57 simulations together with a prescribed ozone dataset, both of which have realised outstanding performances (R(2) > 0.95, RMSE < 2.12 ppbv). The conventional ensemble learning approach is computationally cheaper and results in higher overall performance, but at the expense of oceanic ozone being overestimated and the learning process being uninterpretable. The Bayesian approach performs better in spatial generalisation and enables perceivable interpretability, but induces heavier computational burdens. Both of these multi-stage machine learning-based approaches provide frameworks for improving the fidelity of composition-climate model outputs for uses in future impact studies.