MACHINE LEARNING APPLICATIONS IN STRUCTURAL DYNAMIC RESPONSE PREDICTION OF HIGH-RISE BUILDINGS UNDER WIND AND SEISMIC EXCITATIONS: A SYSTEMATIC REVIEW

Abstract

The prediction of structural dynamic responses in high-rise buildings under wind and seismic excitations is essential for performance-based engineering, yet traditional physics-based methods remain computationally prohibitive for real-time and large-scale applications. Machine learning has emerged as a promising alternative, but the field lacks a coherent synthesis of its diverse methodological developments and practical limitations. We conducted a systematic review following the PRISMA framework to critically map and evaluate the existing evidence on machine learning applications for this problem. Our objectives were to categorize the predominant ML architectures, identify critical gaps in validation and generalizability, and assess reported effectiveness in terms of accuracy, efficiency, and robustness. A comprehensive search across seven major databases was performed using the Population-Exposure-Outcome framework, with inclusion criteria requiring quantitative validation against physical data or high-fidelity simulations. The review reveals a rapidly expanding field where recurrent neural networks, particularly Long Short-Term Memory networks, dominate time-series prediction tasks and generally outperform feed-forward alternatives by capturing temporal dependencies. A notable methodological shift is the integration of physics-informed neural networks, which embed governing equations into the loss function and thereby improve predictive accuracy and physical plausibility, especially under data scarcity. Transfer learning has also gained traction for adapting models across building typologies and hazard scenarios. However, key challenges persist: data scarcity for extreme events, poor generalizability across diverse structural configurations, and the absence of standardized error metrics preclude meaningful cross-study comparison. We conclude that machine learning provides a computationally efficient paradigm for dynamic response prediction, with hybrid data-physics approaches showing the most promise for overcoming current limitations. Nevertheless, the gap between proof-of-concept studies and validated, generalizable tools remain substantial, primarily due to the lack of standardized development and validation frameworks. Future efforts should prioritize open benchmark datasets and performance metrics that reflect both accuracy and practical decision-making utility.