A Reliable Non-tuned Machine Learning Approach for Local Scouring Simulation Around Twin Bridge Piers

In the current research, the “Extreme Learning Machine (ELM)” model as an optimized machine learning approach was developed for first time in order to estimate the scour forming around twin bridge piers. Basically, the measurement of the depth of the scour occurring around bridge piers has always been known as an influencing variable in the design of such structures. Factors affecting the depth of the scour gap forming around twin bridge piers were detected. After that, regarding the input variables, four different ELM models were produced. As an another objective of this research, i.e., evaluating the modeling accuracy, the Monte Carlo simulations were implemented and the fivefold cross validation technique was also utilized for validating. Then, among all activation functions, sin was determined as the most optimal one. In the next step, the outcomes of all ELMs were tested and the best one was detected. In addition, the results of the ELMs were placed against the outcomes acquired from the artificial neural network (ANN) to reveal that the former estimates scour amounts with higher exactness. As an example, the estimations of R 2 and the scatter index for the most powerful ELM model were individually calculated to be 0.975 and 0.084. As the most noticeable conclusion of the sensitivity analysis, it was revealed that the Froude number was the most influencing variable. Besides, an uncertainty analysis was executed to display that the ELM model performs overestimated. In addition, a relationship was proposed with the ability of reproducing the depth of the scour occurring around twin bridge piers via the superior model and as the final stage of the research procedure, a partial derivative sensitivity analysis (PDSA) was also performed on all input variables.