PREDICTING THE MECHANICAL PROPERTIES OF COAL-BEARING ROCKS FROM SIMPLE TEST RESULTS USING MACHINE LEARNING ALGORITHMS

İskender Y., Karaman K., Fereıdoonı D., Şahinoğlu E., Cihangir F.

Türkiye 24. Uluslararası Kömür Kongresi ve Sergisi, Zonguldak, Türkiye, 23 - 25 Eylül 2026, cilt.1, ss.97-113, (Tam Metin Bildiri)

  • Yayın Türü: Bildiri / Tam Metin Bildiri
  • Cilt numarası: 1
  • Basıldığı Şehir: Zonguldak
  • Basıldığı Ülke: Türkiye
  • Sayfa Sayıları: ss.97-113
  • Karadeniz Teknik Üniversitesi Adresli: Evet

Özet

This study investigates the use of machine learning (ML) techniques to predict the mechanical properties of coal-bearing rocks, including Uniaxial Compressive Strength (UCS), Brazilian Tensile Strength (BTS), and Point Load Index (PLI), using non-destructive physical index parameters. Accurate determination of these properties is critical for safe and economical geotechnical engineering; however, conventional laboratory testing is often limited by the difficulty of obtaining intact rock cores from weak or highly fractured formations. To address this challenge, three ML algorithms—Linear Regression (LR), K-Nearest Neighbors (kNN), and Artificial Neural Networks (ANN)—were evaluated using dry unit weight (γd), effective porosity (ne), and ultrasonic P-wave velocity (UPV) as input variables. Experimental data were collected from coal and host rock samples obtained from the Zonguldak hard coal field in Türkiye. Comparative analysis revealed that the kNN algorithm consistently achieved the highest predictive performance, particularly when UPV was used as the predictor. The best kNN models produced coefficient of determination (R²) values of 0.832 for BTS, 0.760 for PLI, and 0.822 for UCS, while also yielding the lowest prediction errors, with RMSE values of 1.727 MPa, 0.658 MPa, and 10.397 MPa, respectively. In contrast, models based solely on effective porosity demonstrated poor predictive capability, with R² values below 0.217 and Mean Absolute Percentage Error (MAPE) values exceeding 260%. Sensitivity analysis identified UPV and dry unit weight as the most influential variables, whereas effective porosity contributed minimally. The findings demonstrate that ML, particularly kNN combined with UPV, provides a reliable, cost-effective, and practical alternative for indirect strength estimation of coal-bearing rocks when destructive testing is impractical.