AI-Driven Digital Twin Framework for Predictive Risk Management and Resource Optimization in Large-Scale Civil Infrastructure Projects

Abstract

This study investigated an AI-driven digital twin framework designed to enhance predictive risk management and resource optimization in large-scale civil infrastructure projects. A quantitative, data-driven research design was applied to 134 active infrastructure projects spanning transportation, water and wastewater, vertical building, energy, and rail sectors; after data screening for incomplete sensor and schedule records, 121 projects were retained for final analysis. Continuous data streams, including Internet of Things (IoT) sensor readings, Building Information Modeling (BIM) parameters, schedule variance indicators, and labor and equipment utilization rates, were synchronized within project-specific digital twin environments across full construction phases. Multiple regression analysis revealed that digital twin-derived exposure variables significantly predicted structured project risk scores, with the overall model explaining 56.2% of variance (R² = 0.562, p < .001). Schedule deviation intensity (β = 0.40, p < .001) and structural sensor anomaly frequency (β = 0.30, p < .001) emerged as the strongest predictors of elevated risk classification. A graph neural network integrated with long short-term memory layers (GNN-LSTM) achieved a risk classification accuracy of 88.9%, outperforming random forest (81.4%), logistic regression (72.6%), and decision tree (68.9%) models, with statistically significant differences across paired comparisons (p < .001). Cumulative exposure analysis indicated a positive correlation between aggregated digital twin risk index and cost overrun magnitude (r = 0.55, p < .001), confirming the framework's relevance to resource optimization outcomes. Reliability testing demonstrated strong internal consistency for the project risk exposure construct (α = 0.87) and the resource optimization index (α = 0.84). Mixed-use mega-projects accumulated the highest proportion of high-risk classification windows at 34.6%, compared to 17.9% for vertical building projects. Findings demonstrate that AI-driven digital twin modeling provides a reliable, objective, and scalable complement to traditional project controls, supporting its application as a predictive decision-support framework for enhancing risk management and resource allocation in large-scale civil infrastructure delivery.