Projecting the Impacts of Climate Change on Ozone and Public Health at High Resolution Using a Machine Learning Method

Description This repository contains the core code and processed input datasets used for the study “Projecting the Impacts of Climate Change on Ozone and Public Health at High Resolution Using a Machine Learning Method.” The repository provides the computational workflow for training and evaluating an XGBoost-based ozone prediction model, applying the trained model to historical and future climate datasets, estimating climate-change-induced changes in surface ozone, assessing the individual contributions of meteorological variables, and evaluating extrapolation risk using an Area of Applicability framework. The workflow is organized into four Jupyter notebooks: 1.Train_xgboost.ipynbThis notebook trains the main XGBoost ozone prediction model using the processed historical dataset. It includes model tuning, validation and test evaluation, spatial and statistical performance diagnostics, feature-importance analysis, and SHAP-based model interpretation. The main input file is: Ozone_prediction_input_v3.4.npy Main outputs include the trained XGBoost model, selected feature list, model metadata, best model configuration, performance figures, residual diagnostics, spatial RMSE maps, and SHAP-related results. 2.Model_evaluation_XGBoost.ipynbThis notebook performs an additional independent model evaluation using 2015–2016 data. It applies the trained XGBoost model to a separate evaluation dataset and produces model performance diagnostics for this independent period. The main input file is: Ozone_prediction_2015_2016_v3.4.npy Main outputs include scatter/hexbin evaluation plots and performance metrics for the independent 2015–2016 evaluation period. 3.Model_prediction_XGBoost.ipynbThis notebook applies the trained XGBoost model to baseline and future climate datasets to estimate climate-change-induced ozone changes. It predicts annual and seasonal ozone changes under future climate scenarios and evaluates the individual effects of meteorological variables by using one-at-a-time future meteorological perturbation datasets. The main baseline input file is: Ozone_prediction_input_2001-2010_v3.4.npy Future prediction datasets are stored in: Prediction_Dataset/ One-at-a-time meteorological perturbation datasets are stored in folders such as: Prediction_Dataset_Only_DTR/ Prediction_Dataset_Only_hursmax/ Prediction_Dataset_Only_hursmin/ Prediction_Dataset_Only_huss/ Prediction_Dataset_Only_pr/ Prediction_Dataset_Only_rsds/ Prediction_Dataset_Only_tasmax/ Prediction_Dataset_Only_tasmin/ Prediction_Dataset_Only_uas/ Prediction_Dataset_Only_vas/ Prediction_Dataset_Only_wspeed/ Main outputs include gridded ozone predictions, absolute and percentage ozone changes relative to the 2001–2010 baseline, annual and seasonal ozone-change maps, county-level model output files, and diagnostic summaries for future climate scenarios. 4.AoA_Estimation.ipynbThis notebook estimates extrapolation risk for future projections using an Area of Applicability approach. It compares future predictor-space conditions with the historical training domain and summarizes the fraction of future grid-month samples that fall outside the historical applicability range. The main inputs are: Ozone_prediction_input_v3.4.npy future prediction datasets in Prediction_Dataset/ Main outputs include extrapolation-risk summary tables and figures for annual, summer, and winter conditions across future scenarios. Repository structure The repository includes the following core notebooks: 1.Train_xgboost.ipynb 2.Model_evaluation_XGBoost.ipynb 3.Model_prediction_XGBoost.ipynb 4.AoA_Estimation.ipynb The processed input datasets are provided as NumPy .npy files. These files contain the gridded ozone, meteorological, land-use, geographic, population, and emissions-related variables required by the notebooks. Future prediction datasets are organized by scenario and by meteorological-variable perturbation experiment. Input datasets The main input datasets are: Ozone_prediction_input_v3.4.npyHistorical dataset used for model training, validation, testing, SHAP analysis, and Area of Applicability estimation. Ozone_prediction_2015_2016_v3.4.npyIndependent 2015–2016 dataset used for additional model evaluation. Ozone_prediction_input_2001-2010_v3.4.npyBaseline dataset used for historical reference predictions in the future ozone-change analysis. Future scenario datasets in Prediction_Dataset/Processed future climate datasets used to estimate total climate-change-induced ozone changes. Future one-at-a-time meteorological perturbation datasets in Prediction_Dataset_Only_* foldersProcessed datasets used to isolate the individual effects of selected meteorological variables on future ozone projections. Outputs Running the notebooks produces: Trained XGBoost model files Selected feature lists and model metadata Model tuning and performance diagnostics Validation, test, and independent-evaluation figures SHAP feature-importance and model-interpretation results Annual and seasonal ozone predictions Absolute and percentage ozone-change maps County-level model output files for health-impact analysis One-at-a-time meteorological contribution estimates Area of Applicability and extrapolation-risk summaries Suggested workflow The notebooks are intended to be run in the following order: Run 1.Train_xgboost.ipynb to train the XGBoost model and generate model artifacts. Run 2.Model_evaluation_XGBoost.ipynb to evaluate the trained model against the independent 2015–2016 dataset. Run 3.Model_prediction_XGBoost.ipynb to generate annual and seasonal ozone projections and meteorological-variable contribution estimates. Run 4.AoA_Estimation.ipynb to quantify extrapolation risk in future predictor space. Users may need to modify file paths in the notebooks depending on their local or high-performance-computing environment. Notes This repository is intended to support reproducibility of the main machine-learning modeling, prediction, and extrapolation-risk analyses reported in the associated manuscript. The provided processed datasets are designed to be directly used by the notebooks without requiring users to repeat the full raw-data preprocessing workflow.