DEMILADE07/oil-presence-prediction

Predicting oil presence from geological and seismic data; SPE DSEATS Africa 2026 Datathon # Prediction of Oil Presence Using Machine Learning My solution for the **SPE DSEATS Africa 2026 Datathon** — predicting whether an unexplored location holds oil (1) or not (0) from geological and seismic measurements. The core idea is to treat this as a **petroleum-system** problem: oil only accumulates when a good reservoir rock, a real trap, a nearby charge, and a supporting seismic indicator all line up in the same place. I show from the data that discovery really does follow this "all at once" rule, then build that logic into the features and the data cleaning. ## What's here | File | Description | |---|---| | `TeamName_PythonCode.ipynb` | The full notebook: data audit, EDA, Part 1 and Part 2, predictions | | `TeamName_Prediction_Part1.csv` | Test-set predictions from the standard workflow | | `TeamName_Prediction_Part2.csv` | Test-set predictions from the physics-informed workflow | | `PROJECT_WALKTHROUGH.md` | A plain-English, beginner-friendly guide to the whole project | | `SOLUTION_OVERVIEW.md` | Approach summary, results, and submission checklist | | `SLIDE_DECK_OUTLINE.md` | 10-slide presentation plan | | `build_notebook.py` | Script that assembles the notebook | | `requirements.txt` | Exact library versions | > The competition datasets and the guidelines PDF are not included here, as they were provided by > the organizers to registered participants. ## Approach in brief - **Part 1 (standard ML):** reconcile the `Trap_Type` encoding, impute missing values, encode categories, and compare six models (Logistic Regression, Random Forest, Extra Trees, HistGradientBoosting, XGBoost, LightGBM) with 5-fold cross-validation, then average the best. - **Part 2 (physics-informed):** correct five geological inconsistencies across `Trap_Type`, `Porosity` and `Permeability`, impute using rock physics, and add petroleum-system features (Reservoir Quality Index, Flow Zone Indicator, and a Chance-of-Success score). ## Results (5-fold cross-validation) About **0.8 …