Optimizing post-earthquake bridge network restoration planning using Graph Neural Networks and Deep Reinforcement Learning

Transportation networks are essential to the regional built environment and are highly vulnerable to seismic hazard impact.Efficient and rapid restoration of bridges following an earthquake is crucial to minimizing disruption and ensuring rapid recovery of the network.However, it is a challenging spatial-temporal decision-making problem, where sequential repair decisions should be made to the spatially distributed bridges within the transportation network under various uncertainties and constraints.This study proposes an innovative approach to optimize the restoration planning of post-earthquake bridge networks, by introducing Graph Neural Networks (GNNs) into Deep Reinforcement Learning (DRL) based decision policies.The integrated restoration decision-support tool can holistically consider the bridge network's spatial and topological information and the complex temporal dynamics during the restoration process.The restoration decisions are optimized under budget constraints through a compound reward function design that aggregates multiple reward signals during the restoration process, such as indirect costs, budget exceedance, and incentives for repair actions.The proposed GNN-DRL agent can effectively coordinate and prioritize network-level bridge repairs based on the bridges' damage states, restoration status, connectivity, and importance within the network.This method is evaluated on two small-scale bridge networks, with a comparative study against a ranking based benchmark approach.The results demonstrate that the integrated GNN-DRL agent outperforms the ranking based approach, achieving a higher success rate and improved restoration performance.The findings reflect promising results in the effectiveness of combining GNN and DRL for efficient decision-making of spatial bridge networks, ensuring the restoration of high-priority bridges and minimizing the overall post-earthquake socioeconomic impact.