{
  "best_parameters": {
    "colsample_bytree": 0.7449721611208636,
    "learning_rate": 0.020824306666550746,
    "max_depth": 3,
    "min_child_weight": 3.2773123153428503,
    "n_estimators": 300,
    "reg_alpha": 0.8287375091519293,
    "reg_lambda": 1.1369027867815908,
    "subsample": 0.7860370513913187
  },
  "categorical_features": [
    "degradation_condition",
    "degradation_mechanism",
    "additives",
    "degradation_environment",
    "additive_type_1",
    "additive_type_2",
    "additive_type_3",
    "sample_shape_morphology",
    "pha_degrading_microbes",
    "experimental_scale"
  ],
  "categorical_levels": {
    "additive_type_1": [
      "crosslinker_polymer_modifier",
      "filler",
      "not_applicable",
      "plasticizer",
      "reinforcement"
    ],
    "additive_type_2": [
      "filler",
      "not_applicable",
      "nucleating_agent",
      "plasticizer"
    ],
    "additive_type_3": [
      "compatibilizer",
      "crosslinker",
      "filler",
      "not_applicable",
      "reinforcement"
    ],
    "additives": [
      "no",
      "yes"
    ],
    "degradation_condition": [
      "aerobic",
      "anaerobic"
    ],
    "degradation_environment": [
      "compost_organic_fertilizer",
      "laboratory_mineral_media",
      "marine_aquatic",
      "soil"
    ],
    "degradation_mechanism": [
      "hydrolysis",
      "microbial_enzymatic",
      "other",
      "surface"
    ],
    "experimental_scale": [
      "lab",
      "large"
    ],
    "pha_degrading_microbes": [
      "bacteria_dominated",
      "bacteria_fungi_mixed",
      "diverse_environmental",
      "fungi_dominated",
      "uncategorized"
    ],
    "sample_shape_morphology": [
      "films",
      "other",
      "ribbons",
      "sheets"
    ]
  },
  "citation": {
    "companion_manuscript": "Bibliographic record pending author submission/publication.",
    "source_corpus": "Kotzabasaki et al. (2026), Polymers 18, 897",
    "source_doi": "10.3390/polym18070897"
  },
  "descriptive_interval_half_widths": {
    "0.80": 53.882681316641424,
    "0.90": 77.78028288264014,
    "0.95": 85.50717191428235
  },
  "endpoint": "degradation-induced weight loss / physical mass-loss trajectory (%)",
  "features": [
    "degradation_time_days",
    "adjusted_hb_ratio_formulation_mol",
    "adjusted_hv_ratio_formulation_mol",
    "t_deg",
    "additive1_percentage_wt",
    "additive2_percentage_wt",
    "additive3_percentage_wt",
    "degradation_condition",
    "degradation_mechanism",
    "additives",
    "degradation_environment",
    "additive_type_1",
    "additive_type_2",
    "additive_type_3",
    "sample_shape_morphology",
    "pha_degrading_microbes",
    "experimental_scale"
  ],
  "grouped_cv_macro_study_mae_objective": 27.81102157801946,
  "inherited_preprocessing": {
    "other": [
      "source-level category coarsening",
      "zero filling of unreported additive concentrations"
    ],
    "temperature_completed_rows": 210,
    "temperature_completed_studies": 3
  },
  "intended_use_limits": "Research hypothesis exploration and experiment prioritisation only; not for certification, regulatory or safety decisions, causal inference, or replacement of laboratory testing. Marginal checks do not validate joint scenarios.",
  "interval_evaluation": {
    "0.80": {
      "empirical_coverage": 0.9284253578732108,
      "mean_width": 79.00438994817615,
      "median_width": 79.76067915555889
    },
    "0.90": {
      "empirical_coverage": 0.9761417859577368,
      "mean_width": 91.92065079330042,
      "median_width": 97.01415518198472
    },
    "0.95": {
      "empirical_coverage": 0.9904567143830948,
      "mean_width": 97.05151882369664,
      "median_width": 100.0
    }
  },
  "interval_method": "Post-selection study-grouped cross-validation residual diagnostic; residual-producing folds also informed hyperparameter selection.",
  "license_scope": "MIT covers project-authored interface code and packaging only; source article, workbook, CSV, and notebooks retain their original rights and are not uploaded.",
  "limitations": [
    "The effective validation sample contains only 16 source studies.",
    "Leave-one-study-out errors are large and heterogeneous.",
    "Prediction intervals are often close to the full 0-100% range.",
    "Inputs outside recorded ranges or categories require caution.",
    "Predictions are associations from literature-derived data, not causal estimates."
  ],
  "model_export_id": "phbv-bm-xgb-2026-07-26-v1",
  "model_name": "bounded_monotonic_study_weighted",
  "monotonic_constraint": {
    "direction": "increasing",
    "feature": "degradation_time_days"
  },
  "nested_loso_performance": {
    "macro_curve_mae": 24.15217559047409,
    "macro_study_mae": 23.5543495445916,
    "point_mae": 22.061178747043066,
    "point_r2": 0.1452965158575815,
    "point_rmse": 30.482867958078305
  },
  "not_endpoint": "ultimate biodegradation or mineralization",
  "numeric_features": [
    "degradation_time_days",
    "adjusted_hb_ratio_formulation_mol",
    "adjusted_hv_ratio_formulation_mol",
    "t_deg",
    "additive1_percentage_wt",
    "additive2_percentage_wt",
    "additive3_percentage_wt"
  ],
  "numeric_ranges": {
    "additive1_percentage_wt": {
      "maximum": 70.0,
      "minimum": 0.0
    },
    "additive2_percentage_wt": {
      "maximum": 50.0,
      "minimum": 0.0
    },
    "additive3_percentage_wt": {
      "maximum": 20.0,
      "minimum": 0.0
    },
    "adjusted_hb_ratio_formulation_mol": {
      "maximum": 99.0,
      "minimum": 28.16
    },
    "adjusted_hv_ratio_formulation_mol": {
      "maximum": 19.0,
      "minimum": 0.5
    },
    "degradation_time_days": {
      "maximum": 360.0,
      "minimum": 0.0
    },
    "t_deg": {
      "maximum": 60.0,
      "minimum": 20.0
    }
  },
  "random_seed": 42,
  "release_date": "2026-07-26",
  "sample_weighting": "study_balanced",
  "software": {
    "numpy": "1.26.4",
    "optuna": "3.2.0",
    "pandas": "2.1.4",
    "python": "3.10.14",
    "scikit_learn": "1.7.2",
    "xgboost": "3.0.5"
  },
  "target_transform": {
    "epsilon": 0.001,
    "limits": [
      0,
      100
    ],
    "prediction_output_clipping": "0-100%",
    "training_target_clip_count": 3,
    "type": "bounded_logit"
  },
  "training_curves": 129,
  "training_population": "Released additive-enriched PHBV modelling corpus: 1,467 observations, 129 curves, and 16 source studies; not a probability sample of PHBV experiments.",
  "training_rows": 1467,
  "training_studies": 16
}