Probabilistic lithofacies classification with machine learning application for improved confidence in subsurface modelling

Abstract

Classifying lithofacies is essential for understanding geological variability in hydrocarbon reservoirs. Machine learning (ML) tools are increasingly used to classify lithofacies for reservoir characterization. In this study, we implemented three ML classification algorithms: Multilayer Perceptron (MLP), Random Forest (RF), and CNN-Shift Padding model to output each individual probabilistic facies class in a range of 0 to 1 using an open-source dataset from the Hugoton and Panoma gas fields, Southwest Kansas, USA. The dataset consists of 4,149 samples from 10 wells with facies identified from core description where each well is characterized by five wireline logs and two derived geological properties. Additional input features and missing data was generated as part of data quality control and data conditioning. To better reflect geological continuity, predictions were evaluated using both exact facies labels and geologically reasonable adjacent facies labels. While RF and MLP yielded comparable predictive metrices, the performance of MLP model was found optimal, achieving higher blind well prediction accuracies when adjacent facies are considered. Incorporation of probability-based facies prediction helped to quantify model confidence and to identify ambiguous zones, particularly within transitional intervals or thin beds. Results show that MLP model reliably detects thin layers of PhylloidAlgal Bafflestone (BS), a critical gas-bearing facies with higher porosity and permeability, even with limited training samples. Overall, the study demonstrates that integrating additional input features, adjacent facies adjustment, and probability thresholding provides a robust framework for lithofacies prediction. The approach improves accuracy beyond previously reported benchmarks and enhances interpretability by quantifying prediction confidence, thereby supporting more reliable geological modelling and reservoir characterization in wells lacking core samples.