Model Checkpoints, Inference Outputs and Plots for Transferable Hourly Ozone Forecasting

This repository contains the complete set of model outputs, trained model checkpoints, training logs, and inference datasets used to generate all results presented in the manuscript. The contents are organized into three main components: All_Plots, checkpoint_n_lightninglogs, and Inf_Data, enabling full traceability from raw inputs to final figures.

1. All_Plots

The All_Plots/ directory contains all figures generated for model evaluation, ablation studies, operational comparison, and transfer learning experiments. The structure is organized by experiment type, geographical domain, and evaluation metric.

• Ablation_Studies/:
  Includes controlled experiments on model components (e.g., context window length, removal of future inputs, exclusion of anthropogenic metadata). Results are stratified by station type (rural, sub-urban, urban) and reported across standard metrics (MAE, RMSE, SMAPE), including type-averaged summaries.
• cams_forecast_comparison/:
  Provides comparisons against Copernicus Atmosphere Monitoring Service (CAMS) forecasts for multiple training configurations. Includes sample-wise predictions, error distributions, and skill score analyses for both training-domain and unseen inference periods.
• ForecastMapPlots/:
  Contains spatial forecast visualizations (e.g., 96-hour horizons and seasonal aggregates) for qualitative assessment of spatial prediction patterns.
• Korea/:
  Documents transfer learning results for South Korea, including sample-level predictions and aggregated performance metrics.
• SkillPlotsWithUncertainty/:
  Includes probabilistic evaluation outputs and uncertainty-aware skill metrics used to assess calibration and forecast reliability.
• TensorBoard/:
  Contains training diagnostics across experiments, including pretraining and transfer learning runs, with metrics tracked across epochs.

2. checkpoint_n_lightninglogs

The checkpoint_n_lightninglogs/ directory contains all trained model checkpoints and associated PyTorch Lightning logs required to reproduce the experiments.

• ablation_studies/:
  Checkpoints corresponding to ablation configurations (e.g., context window variations, no_future, no_meta), with multiple training runs tracked via Lightning versioning.
• pretraining_direct_DE/:
  Pre-trained models on the Germany dataset across different training durations and configurations, including multiple experimental runs and checkpoints.
• TransLearn/:
  Transfer learning experiments for South Korea, including:
  • frozen and unfrozen training strategies,
  • learning rate tuning (including Optuna-based searches),
  • multiple experimental runs tracked via Lightning versions.

Each experiment directory contains:

• checkpoints/: saved model weights for specific runs
• lightning_logs/: training logs, metrics, and run metadata
• station-level metadata (e.g., station type, code, area) where applicable

This structure ensures reproducibility of both model training and evaluation workflows.

3. Inference_Data

The Inference_Data/ directory contains all input data and corresponding model outputs used during inference for both trained model and transfer learning experiments presented in the manuscript.

• Raw test inputs: preprocessed input features used for model evaluation
• Model outputs: predicted ozone concentrations corresponding to each experiment and configuration
• Organized by experiment type, region (Germany / South Korea), and inference period
  
  Note: Inference outputs of ablation studies presented in appendices and additional experiements are not provided here due to size limitations but available on request for specific experiments that are of interest in future. However all corresponding checkpoints of model used including appendix results and source code for downloading data are provided readily in above folders should you be interested to access them.

This directory enables direct verification of model predictions and reproduction of all evaluation metrics and plots presented in the manuscript.