Abstract
Earthquake loading is inherently stochastic and complex, rendering accurate seismic response assessment a computationally demanding. This study employs established machine learning (ML) algorithms for the seismic performance assessment of torsionally irregular reinforced concrete (RC) buildings founded on nonlinear soil conditions, with explicit consideration of soil–structure interaction (SSI) effects. Six building configurations representing varying degrees of plan irregularity, four base conditions encompassing fixed and flexible foundation scenarios, and twenty-two ground motion records were systematically considered. Incremental Dynamic Analysis (IDA) was performed, yielding 3,696 full three-dimensional nonlinear finite element (FE) simulations. The extracted structural response parameters were used to train and validate ML models and to develop GUI-based prediction framework. To assess the generalizability of the model across broader seismic scenarios, 120,000 additional simulations were generated using randomized seismic attributes. The results demonstrate that nonlinear SSI significantly amplifies seismic demand, with response increasing progressively from stiff to soft soil conditions and from regular to torsionally irregular configurations, thereby highlighting the limitations of conventional fixed-base assumptions. Among the evaluated algorithms, XGBoost and LightGBM consistently exhibited the superior predictive performance, accurately capturing the complex nonlinear structural behaviour of the FE models with mean absolute percentage errors below 0.2% and coefficients of determination exceeding 0.976 across all cases. Similarly, feature importance analyses identified Peak Ground Acceleration (PGA) and other dynamic ground-motion characteristics as the dominant parameters governing structural seismic response. Rotational angle as a torsional irregularity measure emerged as a highly sensitive response parameter, exhibiting strong alignment with plan irregularity and SSI effects, which are not fully represented in conventional threshold-based classifications. The developed framework incorporates a comprehensive set of structural, geotechnical, and seismic features, enabling rapid response estimation across a wide range of realistic scenarios. The proposed ML models accurately replicated nonlinear FE simulations with negligible computational cost, demonstrating their effectiveness as high-fidelity surrogate tools for reliable seismic performance assessment under nonlinear SSI conditions.

This publication has 48 references indexed in Scilit: