KLabsAnalytics/kaduna_ozone_ope_project

# Kaduna Ozone Production Efficiency — Kinetic Modelling, Statistical Analysis & ML Surrogate Atmospheric kinetic modelling of ground-level ozone over the Kaduna region, Nigeria (9.75–11.25°N, 6.75–8.25°E) for the period **2015–2024** (120 months). Built on Copernicus CDS (ERA5) and ADS (CAMS EAC4 + AOD550) reanalysis data. This repository covers the full workflow: data integration and EDA → statistical hypothesis testing and sensitivity analysis → ML surrogate model for Ozone Production Efficiency (OPE) with SHAP interpretability → scientific figures → deployed Streamlit prediction app. --- ## Table of Contents 1. Project Status / Scope 2. Repository Structure 3. Reproducing the Analysis 4. Modeling Summary 5. Key Statistical Findings 6. Deployment 7. Data Sources 8. License / Attribution --- ## Project Status / Scope | # | Decision | Justification | |---|----------|---------------| | 1 | **Data coverage: 2015–2024 (120 months)** | All analyses use the full available range. | | 2 | **LSTM replaced by feedforward MLP** | 120-month panel data shows no significant autocorrelative trend — recurrent architectures are poorly suited to this sample. | | 3 | **Full PINN replaced by physics-guided XGBoost** | The underlying relationship is a closed-form algebraic box model, not a PDE. There is no differential-equation residual for a PINN loss term to enforce; explicit Arrhenius/photolysis feature engineering achieves the same physics-informed goal. | --- ## Repository Structure ``` . ├── app/ │ └── app.py # Streamlit deployment app (OPE prediction + SHAP) ├── data/ │ ├── raw/ # Original source files │ │ ├── cds_era5_raw.csv │ │ ├── ads_cams_raw.csv │ │ └── ads_aod550_monthly_aggregated.csv │ └── processed/ # Pipeline outputs (cleaned/combined data, JSON results) │ ├── combined_full.csv # Combined dataset incl. true AOD550 + derived vars │ …