Predicting Stress-Strain Relationships in Stratified Rock Mass by using Machine Learning Techniques

N.S. Melkoumian (Australia)


Stratified rock, stress-strain state, machine learning.


The paper suggests a method based on machine learning techniques to predict the stress-strain relationships in the stratified composite rock mass. To construct the input output dataset for the learning stage, published experimental data for different stratified composite rock specimens have been used. The parameters of the covariance function have been optimized maximizing the log of the marginal likelihood for the experimental results. In the inference stage the obtained parameters are used to statistically predict the stress-strain state of the same rock mass for new stresses not used in the learning stage. The predictions are compared with the actual experimental data aimed to evaluate the validity and applicability of the suggested method. The results demonstrate that by conducting a limited number of experiments and applying the proposed machine learning techniques, it is possible to predict stress-strain relationships for stratified composite rock masses for different combinations of parameters. Application of the proposed method can significantly decrease the number of experiments required for constructing reliable mechanical models for rocks. The uncertainty of the predictions will be high in the case of very limited number of experiments; however it will be significantly reduced once new experimental results are provided to the proposed Bayesian predictive model.

Important Links:

Go Back