{
  "schema_version": 1,
  "experiment_id": "paper_prediction_attempt01",
  "project_id": "ozone_environmental_drivers",
  "paper_doi": "10.1016/j.atmosres.2026.108869",
  "input": "data/03_atmospheric_transport/AQ-Bench/v1/AQbench_dataset.csv",
  "input_sha256": "34bc7d888264ea0d8a37ede2c84d1aab686cc8517e1f77d1c619717a329b1f56",
  "dictionary": "data/03_atmospheric_transport/AQ-Bench/v1/AQbench_variables.csv",
  "dictionary_sha256": "04825047ba63b8d43e574a2a4c5a7637dbf47edfe93e20f674f40c99edc4b59a",
  "id_column": "id",
  "target": "o3_average_values",
  "target_unit": "ppb",
  "split_column": "dataset",
  "development_labels": ["train", "val"],
  "test_label": "test",
  "country_column": "country",
  "minimum_country_count_exclusive": 50,
  "climate_column": "climatic_zone",
  "climate_levels": ["dry", "moist", "wet"],
  "missing_sentinels": [-999],
  "missing_policy": "complete_case_common_features",
  "expected_counts": {"development": 4077, "test": 1009},
  "feature_sets": {
    "model_1": ["climatic_zone", "alt", "water_25km", "grasslands_25km", "croplands_25km", "urban_and_built-up_25km", "no2_column", "population_density", "nightlight_1km"],
    "model_2": ["climatic_zone", "alt", "population_density"]
  },
  "learner": "regr.ranger",
  "folds": 10,
  "candidate_count": 20,
  "num_trees": 500,
  "min_node_sizes": [1, 3, 5, 10, 20],
  "sample_fractions": [0.6, 0.8, 1.0],
  "replace": true,
  "respect_unordered_factors": "partition",
  "selection_metric": "mean_fold_rmse",
  "seed": 20261028,
  "threads": 1,
  "assumptions_not_specified_in_paper": [
    "Merge the released train and val labels into the development cohort; this reproduces the published sample counts.",
    "Count countries before climate exclusions; map terminal dry/moist/wet climate labels and exclude unknown or unmapped labels.",
    "Use a common complete-case cohort for the two feature sets; do not impute targets or predictors.",
    "Use shuffled station-level ten-fold CV shared by both models; no spatial or country blocking.",
    "Search 20 uniformly sampled unique combinations of mtry=1..p, the listed node sizes and sample fractions, with 500 trees and no depth limit.",
    "Use bootstrap sampling with replacement and partition handling of unordered climate factors.",
    "Select by the arithmetic mean of fold RMSE, retaining the first candidate on ties; refit on all development stations.",
    "Use the declared seed and installed package versions; the original seed and exact search space are not available in the article."
  ]
}