Machine Learning Models Predict Lithology with Over 90% Accuracy

Researchers from multiple institutions have published a comparative analysis evaluating the effectiveness of five mainstream machine learning algorithms for predicting lithology from in situ logging data. The study, published in Natural Resources Research in November 2026, addresses the limitations of traditional geological interpretation, which often relies heavily on expert judgment and is associated with significant time costs and uncertainty. By applying data-driven methods, the team aimed to improve the accuracy and efficiency of lithology identification, a critical step in evaluating geological conditions and optimizing subsurface resource exploration for oil and geothermal development.

The research team, led by Ziyu Hao and colleagues, selected five specific algorithms for their comparative framework: XGBoost, support vector machine, K-nearest neighbors, random forest, and Bayesian neural network (BNN). To ensure a fair performance comparison across these different models, the researchers optimized the hyperparameters for each algorithm using genetic algorithms.

A key aspect of this study is its focus on datasets from different well types, a variable that existing literature has rarely tested or compared. The authors noted that previous studies often failed to account for conditions associated with imbalanced data, which can significantly affect model reliability.

The results indicate that when a dataset is sufficiently large and the lithology classes are relatively balanced, all five algorithms achieve excellent predictive performance. Specifically, the models demonstrated accuracies exceeding 90% on measured well datasets.

However, the study highlights a significant limitation when dealing with severely imbalanced data. Under these conditions, the overall prediction accuracy of the models declined significantly, ranging from 55.3% to 81.9%. This finding underscores the critical influence of sample distribution balance on model reliability.

The Bayesian neural network (BNN) was noted for its superior capability in risk-sensitive applications. Unlike the other algorithms, the BNN possesses the ability to quantify prediction uncertainty, providing a measure of confidence alongside its predictions.

Despite the challenges posed by imbalanced data, the algorithms maintained strong predictive performance in the specific task of gas well identification. The study reports that the models achieved accuracies between 85.9% and 92.9% for this application.

The paper provides practical guidance for automated lithology identification and algorithm selection in engineering applications. By comparing the advantages and limitations of the five machine learning approaches, the study offers valuable insights for informed decision-making under realistic geological conditions.

Subject of Research: Comparative analysis of machine learning algorithms for lithology prediction

Article Title: Data-Driven Machine Learning Algorithms for Lithology Prediction: A Comparative Analysis

Article References: Hao, Z., Jiang, G., Cao, Z., Li, M., Li, G., Zhang, M., & Bao, T. (2026). Data-Driven Machine Learning Algorithms for Lithology Prediction: A Comparative Analysis. Natural Resources Research. https://doi.org/10.1007/s11053-026-10782-0

Image Credits: AI Generated

DOI: 10.1007/s11053-026-10782-0

Keywords: machine learning, lithology prediction, geological exploration, XGBoost, Bayesian neural network, data imbalance, subsurface resources, algorithm comparison, Data-Driven, Machine, Learning, Algorithms