Indirect estimation of unconfined compressive strength of carbonate rocks using extreme learning machine

Abstract

The unconfined compressive strength (UCS) of rocks, one fundamental parameter, is widely used in geotechnical engineering. Direct determination of the UCS involves expensive, time-consuming and destructive laboratory tests. These tests sometimes are difficult to be prepared for cracked rocks. In this way, indirect estimation of the UCS of rocks is widely discussed for simplicity and non-destructivity. Conventional methods for indirect estimation of the UCS of rocks are based on regression analysis which has poor accuracy or generalization ability. This paper presents the extreme learning machine (ELM) for indirect estimation of the UCS of rocks according to the correlated indexes including the mineral composition (calcite, clay, quartz, opaque minerals and biotile), specific density, dry unit weight, total porosity, effective porosity, slake durability index (fourth cycle), P-wave velocity in dry samples and in the solid part of the sample. The correlation between the UCS of rocks and each related index is studied by linear regression analysis. Based on this, the ELM approach is implemented for estimation of the UCS of rocks by comparison with other neural networks and the support vector machines (SVM). Also, parameter sensitivity is investigated on the predictive performance of the ELM by two target functions. The results turn out that the ELM is advantageous to the mentioned neural networks and the SVM in the estimation of the UCS of rocks. The ELM performs fast and has good generalization ability. It is a potential robust method for approaching complex, nonlinear problems in geotechnical engineering.