<p>Machine learning (ML) has become a transformative tool in earthquake engineering, offering powerful capabilities to model complex nonlinear patterns in seismic data and improve hazard assessment, earthquake forecasting and structural health monitoring. Despite its rapid adoption, the existing literature lacks a comprehensive synthesis that integrates recent developments across key research areas. This review addresses this gap through a dual-method approach, combining a scientometric analysis of global publication trends, leading authors, contributing countries, funding sources, and journals with a systematic evaluation of 89 representative studies focused on ML-based ground-motion models (GMMs), earthquake prediction models (EQPMs) and structural health monitoring (SHM). The scientometric analysis reveals rapid growth in ML applications since 2015, with China and the United States as dominant contributors and a wide range of interdisciplinary journals serving as major publication sources. The systematic review demonstrates that advanced ML techniques (deep neural networks, ensemble learners, SVMs, etc.) dominate recent studies. In GMMs, ML models routinely improve ground-motion intensity predictions over traditional regression. In EQPMs, ML algorithms identify subtle precursory patterns in seismic catalogs. In SHM, vision-based CNNs (e.g. U-Nets) and hybrid CNN–RNN models achieve high-accuracy damage detection and localization. Key advantages of ML include higher predictive accuracy (e.g. matching near-zero residuals of state-of-the-art GMPEs) and flexibility (automatic feature learning from complex seismic data). ML systems also enable efficient, near-real-time inference (e.g. orders-of-magnitude faster aftershock forecasting). However, challenges remain: seismic datasets are often limited, noisy, and imbalanced, leading to overfitting and limited generalizability. Moreover, complex ML models lack transparency, their interpretability requires developing explainability techniques and rigorous validation to ensure trust. Looking ahead, the review emphasizes the development of explainable, uncertainty-aware ML models, integration of physics-based constraints (e.g. physics-informed neural nets), and establishment of standardized benchmark datasets and evaluation protocols. These steps, along with expanded data resources (real-time sensor networks, satellite measurements, and open-source initiatives), will support robust, transparent, and reproducible ML tools for seismic hazard assessment and resilient infrastructure design.</p>

Machine Learning Innovations in Revolutionizing Earthquake Engineering: A Review

期刊论文
发表日期：

2025年7月30日

查看全文

Umar Ahmad Noor

摘要

Machine learning (ML) has become a transformative tool in earthquake engineering, offering powerful capabilities to model complex nonlinear patterns in seismic data and improve hazard assessment, earthquake forecasting and structural health monitoring. Despite its rapid adoption, the existing literature lacks a comprehensive synthesis that integrates recent developments across key research areas. This review addresses this gap through a dual-method approach, combining a scientometric analysis of global publication trends, leading authors, contributing countries, funding sources, and journals with a systematic evaluation of 89 representative studies focused on ML-based ground-motion models (GMMs), earthquake prediction models (EQPMs) and structural health monitoring (SHM). The scientometric analysis reveals rapid growth in ML applications since 2015, with China and the United States as dominant contributors and a wide range of interdisciplinary journals serving as major publication sources. The systematic review demonstrates that advanced ML techniques (deep neural networks, ensemble learners, SVMs, etc.) dominate recent studies. In GMMs, ML models routinely improve ground-motion intensity predictions over traditional regression. In EQPMs, ML algorithms identify subtle precursory patterns in seismic catalogs. In SHM, vision-based CNNs (e.g. U-Nets) and hybrid CNN–RNN models achieve high-accuracy damage detection and localization. Key advantages of ML include higher predictive accuracy (e.g. matching near-zero residuals of state-of-the-art GMPEs) and flexibility (automatic feature learning from complex seismic data). ML systems also enable efficient, near-real-time inference (e.g. orders-of-magnitude faster aftershock forecasting). However, challenges remain: seismic datasets are often limited, noisy, and imbalanced, leading to overfitting and limited generalizability. Moreover, complex ML models lack transparency, their interpretability requires developing explainability techniques and rigorous validation to ensure trust. Looking ahead, the review emphasizes the development of explainable, uncertainty-aware ML models, integration of physics-based constraints (e.g. physics-informed neural nets), and establishment of standardized benchmark datasets and evaluation protocols. These steps, along with expanded data resources (real-time sensor networks, satellite measurements, and open-source initiatives), will support robust, transparent, and reproducible ML tools for seismic hazard assessment and resilient infrastructure design.