A surrogate modeling framework for inferring internal variability and forced change in surface ozone

Initial-condition ensembles of chemistry-climate models are useful tools for separating anthropogenic signals in atmospheric composition from the noise of internally-generated climate variability. The noise generated by these ensembles can also be leveraged for risk assessment by quantifying the likelihood of extreme pollution outcomes under the same emissions scenario. However, the high computational cost of chemistry-climate models, especially when they include fully interactive chemistry schemes, limits their use for large-ensemble experiments. Here, we propose an efficient surrogate modeling framework for generating synthetic realizations (i.e. ensemble members) of surface ozone projections that can reproduce the statistics of an initial-condition chemistry-climate model ensemble (the 13-member CESM2-WACCM6 ensemble). For a given emissions scenario, our approach infers internal variability in surface ozone by applying a surrogate trained on a single interactive-chemistry realization to monthly meteorological fields from a large initial-condition ensemble run without interactive chemistry. This avoids the need for simulating multiple interactive-chemistry members and is most appropriate when chemistry feedbacks on meteorology are weak (i.e., when meteorological variability is not substantially altered by interactive chemistry). We compare multiple regression-based surrogate modeling approaches including linear, tree-based, and Gaussian process models and assess trade-offs between local training (separate surrogate fits at each grid cell) and global training (a single surrogate fit to all grid cells).

Among our evaluation metrics, we use the area-weighted root mean square error (RMSE) between the synthetic ensemble and the full-complexity model ensemble statistics, evaluated over populated grid cells, to summarize surrogate skill. We compute the RMSE for the externally forced component (ensemble-mean climatology and linear trend) and for the statistics of the internal variability component computed after removing the ensemble-mean (e.g., standard deviation (SD), 90th percentile (q90), exceedance probability above the 90th percentile (P(>q90)). Locally-trained surrogates reproduce the ensemble-mean climatology and trend with very low error (RMSE ~= 0.14–0.17 ppbv and 0.56–0.77 ppbv per 40 years, respectively), whereas global training exhibits substantially larger errors (RMSE ~= 2.1–7.2 ppbv and 2.4–7.4 ppbv per 40 years), indicating that global training struggles to represent the spatially varying forced response set by the spatial pattern of emissions. For internal variability, the local Gaussian process surrogate best reproduces the spread and tail behavior, achieving the lowest errors in SD (RMSE = 0.61 ppbv), q90 (RMSE = 0.51 ppbv), and P(>q90) (RMSE = 0.025, unitless). Overall, our framework enables the efficient generation of a synthetic initial-condition ensemble for surface ozone that can reproduce both the ensemble-mean response and the statistics of the internal variability at a fraction of the computational cost.