Atmospheric ozone is attracting ever-growing research attention, as i) stratospheric ozone depletion will expose the biosphere to stronger hazardous ultraviolet radiation, which has been emphasised in the Montreal Protocol; ii) ozone as a greenhouse gas can alter the atmospheric thermodynamics; iii) ambient ozone is associated with adverse effects on ecosystem and population health through exposure; and iv) sophisticated photochemical mechanisms make the parametrisation for numerical modelling a chasing-deer challenge (Chapter 1). Original research starts from the Coupled Model Inter-comparison Project Phase 6 (CMIP6), which collates an ensemble of Earth system models to simulate the global surface ozone for the purpose of diagnosis-oriented mutual comparison. The author systematically diagnoses 8 models with supports from a collection of auxiliary features and in situ observations (Chapter 2). Interpreted from the semi-quantitative diagnosis outcomes, geographical incoherence can be ascribed to the erroneous emission inventories; incongruous longitudinal trends result from the disagreements in radical simulation; and central biases are attributable to the high uncertainties in simulating the photolysis rate and reservoir formation. Ensemble-based model diagnosis can point out possible directions for the revision of CMIP6 models. Ozone isopleths describe the non-linear responses of ozone concentrations to changes in precursors, nitrogen oxides and volatile organic compounds (VOCs), and thus are pivotal to the determination of ozone regulation requirements. The author innovatively uses the Community Multiscale Air Quality model with High-order Decoupled Direct Method (CMAQ-HDDM) to simulate surface ozone across China domain in 2017, and simultaneously derive ozone isopleths for individual cities (Chapter 3). Interpreted from the city-level isopleths, densely populated metropolitan agglomerations such as Jing-Jin-Ji, Yangtze River Delta, and Pearl River Delta follow the NOx-saturated regime, indicating NOx controlling will increase ozone. Ambient ozone in eastern China generally follows the VOC-limited regime, suggesting reducing VOCs will more effectively suppress the ozone pollution; while contrarily in western regions. Therefore, city-specific ozone isopleths are instrumental in forming differentiated strategies for ozone abatement of Chinese cities. Settling the cross-model discrepancies to achieve more accurate predictions of surface ozone is an unsolved challenge, and methods that overcome structural biases in models going beyond naïve weighted averaging of multiple models is urgently required. Building on CMIP6, a conventional aggressive ensemble-learning-based algorithm is transplanted, and also a more conservative 2-stage enhanced space-time neural network ensembler is optimised to fuse 57 simulations, both of which have revealed outstanding performances (Chapter 4). The conventional approach is computationally cheaper and achieves slightly higher accuracy, but at the expense of sacrificing the model interpretability and leaving the oceanic overestimation unsolved. The conservative approach performs better in spatial generalisation and enables perceivable interpretability, but requires heavier computational burdens, which is a prior choice in multi-model fusion when computation resources permit. Followed by the multi-model fusion, a space-time Bayesian neural network downscaler has been constructed (Chapter 5) to realise urban-rural distinguished 10 km × 10 km spatial resolution surface ozone prediction with excellent methodological reliability and fair prediction accuracy. Based on the predictions in 8-hour maximum daily average metric, the global rural-site surface ozone were 15.1±7.4 ppb higher than urban sites averaged across 30 historical years, with developing countries being of the most evident differences. The globe-wide urban surface ozone were climbing by 1.9±2.3 ppb per decade, except for the de-creasing trends in eastern United States. On the other hand, the global rural surface ozone tended to remain constant, except for the rising trends in China and India. The novel framework contributes to the deep-learning-driven environmental studies methodologically by providing a brand-new feasible way to realise data fusion and downscaling, which maintains high in-terpretability by conforming to the principles of spatial statistics without compromising the prediction accuracy. Furthermore, the spatial resolved monthly surface ozone dataset with multiple metrics has lain a solid foundation for global health impact studies. To provide more accurate estimations for the risk association between long-term ozone exposure and multi-cause mortalities, the author updates the existing systematic reviews by including recent studies and unifying the exposure metrics (Chapter 6). Cross-metric conversion factors are estimated linearly by decadal observations. A total of 25 studies involving 226,453,067 participants are included in the systematic review. The pooled relative risks associated with each 10 ppb incremental ozone exposure, by mean of the warm-season daily maximum 8-h average metric, are RR=1.014 with 95% confidence interval (CI) ranging 1.009–1.019 for all-cause mortality; 1.025 (95% CI: 1.010–1.040) for respiratory mortality; and 1.019 (95% CI: 1.004–1.035) for cardiovascular mortality. Adjustment for exposure metrics lays a more solid foundation for multi-study meta-analysis. To fill in the research gap, the author has estimated the long-term ozone exposure-associated excess mortalities in urban and rural residents globally during 1990-2019 (Chapter 7), by linking the machine-learning-generated high-resolution surface ozone concentration archives (Chapter 4 and 5) with the most up-to-date meta-analysis pooled relative risks (Chapter 6). Ozone exposure-associated all-cause excess mortality was climbing from 0.92 (95% CI: 0.58 to 1.27) million in 1990 to 1.33 (95% CI: 0.83 to 1.85) million in 2019. Rural excess mortalities kept surpassing the urban deaths over the 30 years. Ozone-attributable cardiovascular mortalities were higher than the respiratory deaths, which have been overlooked in previous studies. Urban-rural environmental inequality finally leads to population health injustice, requiring future pertinent pollution control considerations at policy-level.