Scholarlink

<p>Uniaxial compressive strength (UCS) is a key physical parameter of rock and a fundamental indicator in most of the rock strength criteria. Traditional laboratory tests for UCS are costly, time-consuming, and heavily dependent on specialized equipment. To overcome these limitations, this study proposes methods for predicting rock UCS based on metaheuristic optimization of the multilayer perceptron (MLP) algorithm. A database consisting of 237 samples was established, incorporating five easily measurable input features and the target variable UCS. Six metaheuristic optimization algorithms were utilized to optimize a MLP network, generating six optimized MLP models for rock UCS prediction. The performance of the optimization algorithms and MLP models was evaluated using stratified five-fold cross-validation strategy. Results indicate that the optimal network configuration of ReLU-MLP model consists of three hidden layers with 16 neurons each. L2 regularization on MLP significantly enhances model generalization at the cost of a minor reduction in fitting accuracy. The ReLU-MLP model demonstrates both superior accuracy and efficiency, compared with other ensemble learning methods. The six metaheuristic optimization algorithms outperform the deterministic and probabilistic methods in terms of both accuracy and efficiency, with KOA exhibiting the best overall optimization performance. Consequently, among the metaheuristic optimized MLP models, KOA-MLP demonstrates the best performance in terms of prediction accuracy and stability. SHAP analysis indicates that <i>I</i><sub>s(50)</sub>, SRH, and <i>V</i><sub><i>p</i></sub> positively affect rock UCS, whereas <i>P</i><sub><i>n</i></sub> shows a negative relationship and <i>D</i> exerts only a limited influence. <i>I</i><sub>s(50)</sub>is the most influential input feature affecting UCS, followed by SRH, <i>V</i><sub><i>p</i></sub>, and <i>P</i><sub><i>n</i></sub>, whereas <i>D</i> exhibits the least contribution. The proposed models provide rapid and accurate prediction of rock UCS, particularly for transportation infrastructure in challenging geological settings.</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

Predicting Rock Uniaxial Compressive Strength Via Metaheuristic Optimization of Multilayer Perceptron Algorithm

  • 期刊论文

  • 发表日期：

    2026年3月28日

查看全文
  • Guangtuo Bao,
  • Kai Zhang,
  • Xiangdong Niu,
  • Zelin Zhou

摘要

Uniaxial compressive strength (UCS) is a key physical parameter of rock and a fundamental indicator in most of the rock strength criteria. Traditional laboratory tests for UCS are costly, time-consuming, and heavily dependent on specialized equipment. To overcome these limitations, this study proposes methods for predicting rock UCS based on metaheuristic optimization of the multilayer perceptron (MLP) algorithm. A database consisting of 237 samples was established, incorporating five easily measurable input features and the target variable UCS. Six metaheuristic optimization algorithms were utilized to optimize a MLP network, generating six optimized MLP models for rock UCS prediction. The performance of the optimization algorithms and MLP models was evaluated using stratified five-fold cross-validation strategy. Results indicate that the optimal network configuration of ReLU-MLP model consists of three hidden layers with 16 neurons each. L2 regularization on MLP significantly enhances model generalization at the cost of a minor reduction in fitting accuracy. The ReLU-MLP model demonstrates both superior accuracy and efficiency, compared with other ensemble learning methods. The six metaheuristic optimization algorithms outperform the deterministic and probabilistic methods in terms of both accuracy and efficiency, with KOA exhibiting the best overall optimization performance. Consequently, among the metaheuristic optimized MLP models, KOA-MLP demonstrates the best performance in terms of prediction accuracy and stability. SHAP analysis indicates that Is(50), SRH, and Vp positively affect rock UCS, whereas Pn shows a negative relationship and D exerts only a limited influence. Is(50)is the most influential input feature affecting UCS, followed by SRH, Vp, and Pn, whereas D exhibits the least contribution. The proposed models provide rapid and accurate prediction of rock UCS, particularly for transportation infrastructure in challenging geological settings.