Scholarlink

<p>Multimodal pore geometry is a defining feature of carbonate reservoirs, where depositional and diagenetic processes introduce significant permeability variability. However, core analysis and well testing, the primary sources for permeability data, are often expensive, time-consuming, and spatially limited. This study introduces design of experiment hyperparameter optimization, a novel data-driven approach for predicting reservoir permeability using core and conventional well-log data in conjunction with advanced machine learning (ML) algorithms. Three ML techniques, random forest, extreme gradient boosting, and gradient boosting machine (GBM), were employed and optimized using random search, Bayesian optimization, and the innovative Latin hypercube design (LHD) method for hyperparameter optimization (HPO). The workflow incorporates rigorous data-preprocessing techniques, including outlier detection, missing data imputation, scale correction, and normalization, to ensure data reliability and model robustness. Among the tested methods, the LHD–GBM–HPO model achieved the highest accuracy, with an adjusted <i>R</i><sup>2</sup> of 0.9997 and root mean square error (RMSE) of 0.000063 for training, and adjusted <i>R</i><sup>2</sup> of 0.9531 and RMSE of 0.09988 for validation, outperforming all other models. The novelty of this work lies in integrating design of experiments principles, specifically LHD, with advanced ML hyperparameter optimization to address the complexities of heterogeneous carbonate reservoirs. This hybrid approach not only enhances predictive accuracy but also reduces computational costs, establishing a pioneering methodology for efficient and scalable permeability prediction.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

DoEHypOpt: An Efficient Approach of Hyperparameter Optimization in Machine Learning Using Design of Experiments Applied on Multivariate Permeability Modeling

  • 期刊论文

  • 发表日期：

    2025年11月14日

查看全文
  • Watheq J. Al-Mudhafar,
  • Mohammed A. Abbas

摘要

Multimodal pore geometry is a defining feature of carbonate reservoirs, where depositional and diagenetic processes introduce significant permeability variability. However, core analysis and well testing, the primary sources for permeability data, are often expensive, time-consuming, and spatially limited. This study introduces design of experiment hyperparameter optimization, a novel data-driven approach for predicting reservoir permeability using core and conventional well-log data in conjunction with advanced machine learning (ML) algorithms. Three ML techniques, random forest, extreme gradient boosting, and gradient boosting machine (GBM), were employed and optimized using random search, Bayesian optimization, and the innovative Latin hypercube design (LHD) method for hyperparameter optimization (HPO). The workflow incorporates rigorous data-preprocessing techniques, including outlier detection, missing data imputation, scale correction, and normalization, to ensure data reliability and model robustness. Among the tested methods, the LHD–GBM–HPO model achieved the highest accuracy, with an adjusted R2 of 0.9997 and root mean square error (RMSE) of 0.000063 for training, and adjusted R2 of 0.9531 and RMSE of 0.09988 for validation, outperforming all other models. The novelty of this work lies in integrating design of experiments principles, specifically LHD, with advanced ML hyperparameter optimization to address the complexities of heterogeneous carbonate reservoirs. This hybrid approach not only enhances predictive accuracy but also reduces computational costs, establishing a pioneering methodology for efficient and scalable permeability prediction.