Machine learning models to estimate the elastic modulus of weathered magmatic rocks

Creative Commons License

Ceryan N., Özkat E. C., Korkmaz Can N., Ceryan S.

ENVIRONMENTAL EARTH SCIENCES, cilt.80, sa.12, 2021 (SCI-Expanded) identifier identifier

  • Yayın Türü: Makale / Tam Makale
  • Cilt numarası: 80 Sayı: 12
  • Basım Tarihi: 2021
  • Doi Numarası: 10.1007/s12665-021-09738-9
  • Dergi Adı: ENVIRONMENTAL EARTH SCIENCES
  • Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED)
  • Anahtar Kelimeler: Elastic modulus, Weathered magmatic rocks, Machine learning, Porosity, P-wave velocity, Slake durability index, UNIAXIAL COMPRESSIVE STRENGTH, ARTIFICIAL NEURAL-NETWORK, GAUSSIAN PROCESS REGRESSION, ABSOLUTE ERROR MAE, MECHANICAL-PROPERTIES, DEFORMABILITY PROPERTIES, FUZZY INFERENCE, YOUNGS MODULUS, INTACT ROCKS, ENGINEERING PROPERTIES
  • Recep Tayyip Erdoğan Üniversitesi Adresli: Evet

Özet

In recent years, several soft computing models have been proposed to estimate the elastic modulus of magmatic rocks. However, there are lacks in models that consider the different weathering degrees in determining the elastic modulus of rocks. In the literature, mechanical properties are widely used as inputs in predictive models for weathered rocks; however, there are only a few models that use index properties representing the effect of weathering on magmatic rocks. In this study, support vector regression (SVR) Gaussian process regression (GPR), and artificial neural network (ANN) models were developed to predict the elastic modulus of magmatic rocks with different degrees of weathering. The inputs selected by the best subset regression approach were porosity, P-wave velocity, and slake durability index. Key performance indicators (KPIs) were computed to validate the accuracy of the developed models. In addition to KPIs, Taylor diagrams and regression error characteristic (REC) curves were used to assess the performance of the developed prediction models. In this study, considering the difficulties of expressing the error using only RMSE and MAE, a new performance index (PI), PIMAE, was proposed using normalized MAE instead of normalized RMSE. It was also indicated that PIRMSE and PIMAE should be used together in performance analysis. When considering the Taylor diagram, PIRMSE, and PIMAE, the GPR models performed best, and the SVR model performed the worst in both the training and test periods. Similarly, according to the REC curve in both periods, the performance of the SVR was the worst, while the performance of the ANN model was the best. The PIRMSE and PIMAE values of the GPR model for the test data were 1.3779 and 1.4142, respectively, and they were 1.2567 and 1.4139, respectively, for the ANN model. According to the computed response surfaces, an increase in the P-wave velocity, and a decrease in the porosity increased the elastic modulus. However, changes in slake durability index only had a minor effect on the elastic modulus.