Analysis of dynamic soil-pile-structure interaction and prediction with machine learning

Soil–pile–structure interaction is a complex process in which the superstructure, the foundation–pile system, and the surrounding soil mutually influence each other under dynamic conditions, resulting in a distinct mechanical behavior of the overall system. Within the scope of the Turkish Earthquake Code (TEC) 2018, soil–pile–structure interaction can be analyzed using coupled system and substructuring approaches through methods designated as Method I, II, and III. According to TEC 2018 principles, kinematic and inertial effects may be considered separately or simultaneously in the analyses, and earthquake ground motions are required to be scaled in accordance with prescribed rules prior to interaction analyses. Therefore, compliance with TEC 2018 design provisions is essential to ensure solution validity and structural reliability. In this study, in addition to simple scaling and spectral matching methods, a machine learning–assisted hybrid approach was developed to optimize the advantages of these two methods for scaling earthquake ground motion records. Based on TEC 2018 principles, three-dimensional dynamic soil–pile–structure interaction analyses were conducted using ground motions scaled by simple scaling, spectral matching, and the proposed machine learning–based approach. A high-rise pile-founded reinforced concrete building was designed for the numerical analyses, in which the superstructure was modeled using nonlinear materials and elements, while the soil medium was represented by nonlinear springs. The coupled system model, managed through centralized computer software developed within the scope of this study, was analyzed using nonlinear time-history analyses under dynamic loading conditions. The structural responses and performance of the soil–pile–structure system were comparatively evaluated for the different scaling approaches. In addition, the relationships between seismic ground motion characteristics and soil–pile–structure interaction responses were investigated using a machine learning–based random forest algorithm, through which critical structural and geotechnical demands were predicted within a classification-based framework at acceptable accuracy levels. The results demonstrate that machine learning–based scaling and demand prediction provide an effective and reliable framework for the assessment of nonlinear soil–pile–structure interaction problems.