Assessing the uniaxial compressive strength (UCS) of rocks is fundamental to various civil engineering projects, including slope stability and underground excavation. Traditionally, determining UCS has been time-consuming and labor-intensive, often necessitating laboratory tests after core sampling. Recent advancements, particularly the utilization of Measurement While Drilling (MWD) technology, offer new horizons for UCS estimation by allowing real-time monitoring of drilling parameters. This innovative approach can dramatically reduce costs and time associated with obtaining rock strength parameters.
A recent study published by Xie, Li, and Min (2025) takes strides toward refining UCS predictions by leveraging machine learning models applied to extensive datasets. The research evaluates five machine learning techniques—Multilayer Perceptron (MLP), Support Vector Regression (SVR), Convolutional Neural Networks (CNN), Random Trees (RT), and Long Short-Term Memory (LSTM) networks—with remarkable insights on their relative effectiveness. Key findings reveal Random Trees achieving the best predictive performance, with a root mean square error (RMSE) of 15.851, mean absolute error (MAE) of 4.449, standard deviation of residuals (SDR) of 15.292, and a coefficient of determination (R²) value of 0.959 on the test set.
Support Vector Regression also performed commendably, with results showcasing its utility for UCS predictions across varied geological conditions. The study revealed SVR attaining an RMSE of 21.905 and MAE of 17.962, underscoring its capability for stable performance. These results are particularly significant as they validate machine learning's potential to create adaptive models capable of yielding accurate predictions without the extensive overhead associated with conventional UCS determination methods.
The dataset utilized encompasses 197 experimental datasets gathered from previous literature, covering diverse rock types such as granite, limestone, and sandstone, alongside varying drilling parameters—including thrust force and rotation speed. The comprehensive nature of this dataset serves as the backbone for forming reliable predictive models, addressing the challenge of applying machine learning techniques effectively across various rock types and drilling conditions.
Machine learning centers around creating algorithms capable of learning from data and enhancing performance through exposure to information. This study highlights the transition from traditional empirical models toward intelligent algorithms, which significantly enhances predictive capabilities. Especially noteworthy is Random Trees, which achieved high fitting accuracy within the training data, albeit with some signs of overfitting, making it important to monitor model robustness during application.
Beyond just training and testing, the models were validated against independent datasets, affirming their generalization capability. Random Trees not only led the performance among tested models but also reinforced the feasibility of using machine learning for real-time UCS prediction during drilling operations. Findings indicated positive correlations between drilling parameters and UCS, supporting the hypothesis of MWD data optimization.
The validation phase demonstrated the model's robustness, with Random Trees sustaining its leading position even when faced with independent datasets, where it achieved the highest R² values of 0.887. This consistency across validation processes demonstrates the reliability of the machine learning methods explored.
Implementing these predictive models presents considerable advantages for civil engineering sectors involved with rock mechanics, unlocking pathways for reducing time-intensive procedures linked to UCS determination. Effective machine learning applications reflect broader engineering improvements, potentially revolutionizing methods of rock strength assessment and transforming operational efficiencies.
Despite the promising results, challenges persist. The research acknowledges data limitations, stressing the need for expanded datasets to refine model performance and applicability. Future efforts should devote attention to refining methodologies, integrating new technologies, and preparing the groundwork for field applications of these models, fostering advancements within rock engineering practices.
For practical applications, the RT and SVR models inhabit unique positions of efficiency, merging speed with predictive accuracy, which bodes well for the future of real-time rock behavior assessments. This study offers substantial ground for future exploration, merging artificial intelligence with rock science to facilitate intelligent, automated decisions within engineering fields.