Multi-stage Ensemble-learning-based Model Fusion for Surface Ozone Simulations: A Focus on CMIP6 Models

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.

Cluster‐Enhanced Ensemble Learning for Mapping Global Monthly Surface Ozone From 2003 to 2019

“…Other regions suffering from high ozone pollution are less frequently modeled, such as India and Europe. Whereas, our surface ozone product supports environmental health research with full coverage of global lands at monthly resolution, calibrated by the most recent reliable measurements till 2019, which improves upon previous work with finer temporal resolution (DeLang et al., 2021) and more recent observations for training (Sun & Archibald, 2021; Sun et al., 2021). Fourth, our clustering and GAM method can partially account for the unobserved spatial heterogeneity of surface ozone concentrations over the globe, as demonstrated by an increase of ∼0.07 R 2 and decrease in RMSE compared with using a non‐clustered method (Tables S2 and S3 in Supporting Information ) and also as observed by divergent variable importance at different clusters (Figure S7 in Supporting Information ).…”

“…Compared with existing studies, our study has several advantages. First, we collect the most comprehensive surface ozone observations from different sources than prior studies (DeLang et al., 2021; Sun & Archibald, 2021; Sun et al., 2021), with a ∼15% increase in the total number of observations. Second, we include 89 predictors, which can better handle the complex interactions in surface ozone prediction.…”

“…Since they only provided peak season averaged estimation, aggregated yearly estimates may buffer their performance compared with monthly predictions. Sun and Archibald (2021) used a multi‐stage ensemble machine learning (ML) model to fuse the surface ozone observations with free‐running chemistry‐climate models endorsed by Coupled Model Intercomparison Project Phase 6 (CMIP6) and obtained a monthly global ozone product at 2‐degree resolution during 1990–2014. However, coarse temporal or spatial resolution of these products limits their use for evaluating seasonal and regional differences in ozone concentrations and trends.…”

Long-term ozone exposure associated cause-specific mortality risks with adjusted metrics by cohort studies: A systematic review and meta-analysis

“…There were two major units used to quantify the surface O 3 concentrations, nmol mol -1 or parts per billion by volume mixing ratio (ppbV) more frequently used by atmospheric modelling researchers, 17, 18, 25 and milligram per cubic metre by mass concentration (µg/m 3 ) widely used by public health studies. 12 These two units are interchangeable to each other based on the ideal gas law PV n RT , if the air temperatures (T) and pressures (P) are given, as presented in eqs 1–4.…”

“…Some deficiencies are spotted in the two reviews for long-term O 3 exposure-associated mortality risk studies, 15, 16 the primary issue of which is the inconsistent use of various O 3 exposure metrics; however, no other reviews are found to remedy these flaws. As a secondary photolytic gaseous air pollutant, the warm-season and diurnal concentrations of surface O 3 will be much higher than cool-season and nocturnal concentrations, 17, 18 and thus the average and peak metrics of O 3 concentrations shall be of drastically different realistic implications. 19 Under this circumstance, directly pooling the relative risks scaled in different metrics might lead to biases.…”