Large Language Models Integrated with Digital Twins for Intelligent Structural Engineering Decision Support: A Comprehensive Literature Review

Abstract

This review synthesizes current advances in integrating Large Language Models with Digital Twin technologies for structural engineering applications. LLMs enhance DTs through natural language processing capabilities, enabling intuitive human-machine interfaces for decision-making. The fusion creates intelligent systems capable of processing unstructured data from technical manuals, sensor networks, and historical records. Applications span structural health monitoring, predictive maintenance, design optimization, and risk assessment. Key findings demonstrate that LLM-augmented DTs achieve 96% prediction accuracy in fault detection compared to traditional methods achieving 75%. The integration reduces manual effort by up to 80% in causal model generation and accelerates response times from days to minutes. However, implementation challenges persist including data quality standardization, interoperability issues, and cybersecurity vulnerabilities. The framework emphasizes neuro-symbolic approaches combining neural language processing with symbolic reasoning for trustworthy, transparent autonomous decision-making. This convergence represents a paradigm shift from reactive monitoring toward proactive, adaptive infrastructure management capable of real-time anomaly detection, predictive failure analysis, and autonomous resource optimization. Future research must address model validation, ethical considerations, and cross-disciplinary collaboration.