Petroleum Target Explorer (PTX-AIM): A Rigorously Validated Machine Learning Framework for TOC Prediction in Frontier Basins, a Case Study from the Kolmani River-2 Well, Gongola Basin, Upper Benue Trough, Nigeria
Abstract
Abstract Objectives/Scope The Kolmani River discovery confirmed a working petroleum system in Nigeria's Gongola Basin (Upper Benue Trough), but too few wells and sampled intervals exist for conventional Total Organic Carbon (TOC) prediction to be reliable. We present PTX-AIM (Petroleum Target Explorer – Artificial Intelligence Machine), an open-source framework that predicts TOC from standard wireline logs and compares machine learning models under leakage-free validation. The objective is to test whether machine learning can outperform the industry-standard Passey ΔlogR method where data are scarce. Methods, Procedures, Process The workflow runs end-to-end: LAS/CSV ingest, IQR cleaning, user-selectable feature selection, scaling, and model selection across five algorithms, with an automated machine learning (AutoML) search over the same library. Evaluation uses spatially blocked validation: contiguous 50-ft depth blocks are held out entirely, so no test interval shares a depth neighbour with the training data. This removes the depth-adjacency leakage that inflates reported ML accuracies on single-well data. Results, Observations, Conclusions Under blocked validation on KR-2, a manually tuned Random Forest performed best (R2 = 0.288, RMSE = 0.261 wt%, MAE = 0.182 wt%). Calibrated Passey ΔlogR returned R2 = 0.138, RMSE = 0.287 wt%, MAE = 0.216 wt%. SHAP attribution identified neutron porosity as the dominant predictor. These figures are deliberately conservative: the skill remaining once leakage is removed, not the inflated R2 a random split yields on 0.25-ft data. For a basin with two to three wells, ML extracts more usable TOC signal from existing logs than ΔlogR. Novel/Additive Information This is a first leakage-controlled ML study of TOC prediction on Nigerian frontier-basin data, treating depth-adjacency leakage as a methodological contribution rather than reporting inflated accuracies. It couples user-driven model selection with a natural-language chatbot, making interpretable analysis accessible without ML expertise, and transfers to other frontier basins.
