Prediction of inelastic displacement ratios in soil-structure interaction on very soft soils using neural architecture search-based ML hybrid technique

Abstract

Performance-based seismic design focuses on limiting building lateral inelastic displacements to control potential structural damage during earthquakes. In this study, advanced machine learning methods are used to carry out a new model for predicting inelastic displacement ratios (IDR) in multistorey buildings built on very soft soils, considering soil-structure interaction (SSI) effects. The proposed model enhances prediction accuracy, reduces computational cost, and facilitates real-world seismic response assessments. A comprehensive dataset was generated, encompassing various dynamic characteristics and key SSI parameters of soil-structure systems. Nonlinear time history analyses (NLTHA) were conducted using a set of 20 ground motions recorded on very soft soil sites. The research utilizes artificial neural networks (ANN), random forest (RF) algorithms, and hybrid models optimized via neural architecture search (NAS-ANN and -RF). A practical and user-friendly graphical interface, named "IDRs_SSI2025", has been developed to support the application of the model proposed by engineers and researchers. Results indicate that the proposed methodology improves prediction accuracy, reduces computational cost, and facilitates real-world seismic response assessments.