A comparison of artificial neural network and classical regression models for earthquake-induced slope displacements

Abstract Empirical predictive models of earthquake-induced slope displacements are commonly developed through classical regression techniques that predefine a functional form and derive the model parameters using a large dataset of computed displacements. Machine learning (ML) techniques represent an attractive alternative to classical regression techniques because they can more easily capture complex interactions between the input and output variables without stipulating a specific functional form. This study develops predictive models for seismic slope displacement using the artificial neural network (ANN) approach and a database of slope displacements from finite element analyses, and compares the resulting ANN model with a classical regression model derived from the same dataset. The ANN utilizes the yield acceleration for the slope ( k y ), the natural period of the slope ( T s l o p e ), the depth of the slip surface relative to the height of the slope ( H r a t i o ), and the peak ground velocity (PGV) of the input motion to predict the slope displacement. Relative to the classical regression model generated from the same dataset, the ANN model modestly improves the median displacement at large and small PGV values and reduces the total variability in the displacement prediction by about 5%. The ANN model predicts displacements that vary more smoothly with the input parameters than the functional form used in the classical model. Although the ANN model only moderately improves predictions relative to the classical regression model, as functional forms for classical regression models become more complex, the role of ANN in providing a simpler regression alternative will grow.