Unified Framework for Probabilistic Prediction of Critical Mechanical Parameters of Loess by Machine Learning Methods

In order to predict the criticalmechanical parameters of loess accurately and quantify the uncertainty corresponding to the prediction results reasonably, anunified framework for probabilistic prediction of critical mechanical parameters of loess by machine learning methods is proposed. By fitting probability density function to the bias of the training dataset, a 95% confidence interval for the prediction results is constructed, and the size of the confidence interval reflects the rationality of the prediction results. (Result) Predicting cohesion of loess based on four machine learning methods, namely, random forest, decision tree, extreme gradient boosting and adaptive boosting, the corresponding coefficients of determination R^2 reached 0.84, 0.75, 0.81 and 0.79, respectively. The proportion of measurement data included in the 95% confidence interval constructed by the four methods is around 95%. It is shown that the 95% confidence interval obtained from the bias based on the training dataset is relatively reliable and can quantify the uncertainty of the prediction results reasonably. In addition, the cohesion of loess can be predicted accurately using the four machine learning methods.