Dataset for "Machine Learning Surrogates for Plume Shine Dose Prediction: The Role of Dataset Resolution, Inductive Bias, and Interpretability"

This Zenodo record contains the datasets, trained machine learning models, and supporting preprocessing artifacts used in the study on surrogate modeling of plume shine dose for radiological consequence assessment. The repository includes preprocessed discrete datasets, interpolated continuous datasets, and trained models developed to predict plume shine dose as a function of downwind distance, release height, radionuclide identity, and atmospheric stability category. Interpolation was performed along the downwind distance dimension using shape-preserving methods to enhance spatial resolution while maintaining physical monotonicity of dose attenuation. Contents Trained machine learning models xgboost_model_final_ep100_dep30.json: Trained XGBoost regression model random_forest_final_ep100_dep15.pkl: Trained Random Forest regression model tabnet_final.pkl: Trained TabNet deep learning model Preprocessing and encoding artifacts scaler.pkl: Feature scaling object used during model training label_encoders.pkl: Label encoders for categorical variables (radionuclide and stability category) Datasets filtered_distance_2000_height_200_train_99.csv: Original discrete training dataset filtered_distance_2000_height_200_test_1.csv: Independent test dataset (real, non-interpolated) finer_interpolated_data_filtered_distance_2000_height_200_train_99.csv: Distance-wise interpolated training dataset finer_interpolated_data_filtered_distance_2000_height_200_train_9975.csv: Interpolated dataset with additional augmentation finer_interpolated_data_filtered_distance_2000_height_200_test_0025.csv: Held-out validation dataset Notes Interpolated datasets were generated using physically consistent one-dimensional interpolation along the downwind distance axis, performed separately for each radionuclide–release height–stability category combination. The independent test dataset contains only original (non-interpolated) samples and was not used during model training. The trained models and preprocessing objects are provided to ensure full reproducibility of the reported results.