Application of machine learning in predicting the rate-dependent compressive strength of rocks

Accurate prediction of compressive strength of rocks relies on the rate-dependent behaviors of rocks, and correlation among the geometrical, physical, and mechanical properties of rocks. However, these properties may not be easy to control in laboratory experiments, particularly in dynamic compression experiments. By training three machine learning models based on the support vector machine (SVM), back-propagation neural network (BPNN), and random forest (RF) algorithms, we isolated different input parameters, such as static compressive strength, P-wave velocity, specimen dimension, grain size, bulk density, and strain rate, to identify their importance in the strength prediction. Our results demonstrated that the RF algorithm shows a better performance than the other two algorithms. The strain rate is a key input parameter influencing the performance of these models, while the others (e.g. static compressive strength and P-wave velocity) are less important as their roles can be compensated by alternative parameters. The results also revealed that the effect of specimen dimension on the rock strength can be overshadowed at high strain rates, while the effect on the dynamic increase factor (i.e. the ratio of dynamic to static compressive strength) becomes significant. The dynamic increase factors for different specimen dimensions bifurcate when the strain rate reaches a relatively high value, a clue to improve our understanding of the transitional behaviors of rocks from low to high strain rates.