Predictive Analysis of Slope Stability via Metaheuristic Algorithms Helping Neural Networks

ABSTRACT Attaining a firm slope stability analysis holds eminent importance in civil and geotechnical projects. This study is concerned with the indirect assessment of the stability of slopes using improved versions of artificial neural networks (ANN). Two novel metaheuristic techniques, namely the seeker optimization algorithm (SOA) and the electromagnetic field optimization (EFO) are employed for optimising the ANN that aims at predicting the factor of safety (FOS). Two hybrids of EFO‐ANN and SOA‐ANN, as well as a single conventional ANN, are trained and tested using a valid dataset collected from the earlier literature. First, examining the importance of the input factors showed that the unit weight of slope material ( γ ) is the most important one, followed by internal friction ( ϕ ), average angle of slope ( β ), cohesion ( c ), the height of slope ( H ), and pore water pressure coefficient ( r u ). Upon monitoring the performance of the ANN, this model stops training after some epochs because of the divergence in the solution, whereas this issue was resolved by the EFO and SOA in hybrid models. Consequently, significant improvements were achieved in both training and testing accuracies. By comparison, while the EFO‐ANN was more successful in the training task, the SOA‐ANN presented the most reliable prediction of the FOS. The prediction competency of these models is also verified through (a) comparison with earlier literature and (b) applying them to another real‐world dataset for binary prediction of slope stability/failure. An explicit predictive formula is derived from the SOA‐ANN, which is recommended as a convenient approximator of FOS in slope stability analysis.