Comprehensive knowledge and analysis of in situ rock strength and geo-mechanical characteristics of rocks are crucial in hydrocarbon and mineral exploration stage to maximize wellbore performance, maintain wellbore stability, and optimize hydraulic fracturing process. Due to the high cost of laboratory-based measurements of rock mechanics properties, the log-based prediction is a viable option. Nowadays, the machine learning tools are being used for estimation of the in situ rock properties using wireline log data. This paper proposes a machine learning approach for rock strength (uniaxial compressive strength) prediction. The main objectives are to investigate the performance of data-driven predictive model in determining this vital parameter and to select features of predictor log variables in the model. The backpropagation multilayer perception (MLP) artificial neural network (ANN) with Levenberg–Marquardt training algorithm as well as the least squares support vector machine (LS-SVM) with coupled simulated annealing (CSA) optimization technique is employed to develop the dynamic data-driven models. Capturing nonlinear, high dimensional, and complex nature of real field log data, the rock strength models’ performances are evaluated using statistical criteria to ensure concerning the model reliability and accuracy. The model predictions are compared and validated against the measured values as well as the results obtained from existing log-based correlations. Both the MLP-ANN and the CSA-based LS-SVM connectionist strategies are able to predict the rock strength so that there is a very good match between the model results and corresponding measured values. The input log parameters are ranked based on their contributions in prediction performance. The acoustic travel time and gamma ray are found to have the highest relative significance in estimating rock strength. New correlations are also developed to obtain the in situ rock strength of the siliciclastic sedimentary rocks using the most important log parameters such as dynamic sonic slowness, formation electron density, and shalyness effect. The developed correlations can be used to obtain quick estimation of dynamic uniaxial compressive strength profile using wireline logging data, instead of static data from the surface measurements or laboratory data. It is expected that the proposed models and tools will enable oil and gas engineers to better predict rock strength and thus enhance wellbore performance.

中文翻译：

---

模拟岩石强度的机器学习方法：借助测井数据进行预测和变量选择

对原位岩石强度和岩石地质力学特征的综合知识和分析对于油气勘探阶段至关重要，以最大限度地提高井眼性能、保持井眼稳定性和优化水力压裂工艺。由于基于实验室的岩石力学特性测量成本高昂，基于测井的预测是一个可行的选择。如今，机器学习工具正被用于使用电缆测井数据估计原位岩石特性。本文提出了一种用于岩石强度（单轴抗压强度）预测的机器学习方法。主要目标是研究数据驱动预测模型在确定这一重要参数方面的性能，并选择模型中预测器日志变量的特征。采用具有 Levenberg-Marquardt 训练算法的反向传播多层感知 (MLP) 人工神经网络 (ANN) 以及采用耦合模拟退火 (CSA) 优化技术的最小二乘支持向量机 (LS-SVM) 来开发动态数据——驱动模型。捕捉真实现场测井数据的非线性、高维和复杂性，使用统计标准评估岩石强度模型的性能，以确保模型的可靠性和准确性。将模型预测与测量值以及从现有的基于对数的相关性中获得的结果进行比较和验证。MLP-ANN 和基于 CSA 的 LS-SVM 联结主义策略都能够预测岩石强度，因此模型结果与相应的测量值之间存在非常好的匹配。输入日志参数根据它们对预测性能的贡献进行排名。发现声波传播时间和伽马射线在估计岩石强度中具有最高的相对重要性。还开发了新的相关性，以使用最重要的测井参数（如动态声波时差、地层电子密度和泥质效应）获得硅质碎屑沉积岩的原位岩石强度。开发的相关性可用于使用电缆测井数据快速估计动态单轴抗压强度剖面，而不是来自表面测量或实验室数据的静态数据。预计所提出的模型和工具将使石油和天然气工程师能够更好地预测岩石强度，从而提高井筒性能。