LINEAR VERSUS NON-LINEAR MACHINE LEARNING FEATURE SELECTION FOR EROSION RATE PREDICTION

Machine learning feature has been utilized to demonstrate the importance and prioritize the parameters in erosion rate prediction. With the expansion of experimental data, application of machine learning models has become more feasible. Feature selection is critical for machine learning to make models more physically interpretable while reducing the dimensionality of the models to reduce over fitting. Previous data-driven erosion rate predictions used all parameters from the experimental data to train the models. Thus, previous studies did not examine feature importance and correlation structure during training of machine learning models. In the current study, first a correlation to identify linear relationships between variables and the erosion rates has been used. However, not all the parameters are linearly correlated with the physics of the erosion process. Moreover, extra model parameters in a high dimensional model could lead to over fitting. Thus, we compare linear (Elastic Net, Penalized Linear Regression) and non-linear (Random Forest, Extreme Gradient Boosting) machine learning feature selection for erosion rate prediction to find the most important minimal set of features affecting erosion rate. Understanding the differences between features selected by linear and non-linear methods provides us with experimental guidance and predictive inputs for more physically motivated and generalizable erosion prediction models.