[Paper Review] Graph Neural Network Surrogate for Seismic Reliability Analysis of Highway Bridge Systems

Tong Liu, Hadi Meidani|arXiv (Cornell University)|Oct 12, 2022

Concrete Corrosion and Durability4 citations

TL;DR

This paper proposes a graph neural network (GNN) surrogate model for rapid seismic reliability assessment of highway bridge systems, enabling efficient, accurate, and scalable node-level connectivity prediction across origin-destination pairs. The approach achieves high computational efficiency and robustness by replacing Monte Carlo simulations with an end-to-end GNN framework that supports inductive transfer learning, demonstrating strong performance on real-world California bridge networks with F1 scores above 0.92 even in out-of-distribution settings.

ABSTRACT

Rapid reliability assessment of transportation networks can enhance preparedness, risk mitigation, and response management procedures related to these systems. Network reliability analysis commonly considers network-level performance and does not consider the more detailed node-level responses due to computational cost. In this paper, we propose a rapid seismic reliability assessment approach for bridge networks based on graph neural networks, where node-level connectivities, between points of interest and other nodes, are evaluated under probabilistic seismic scenarios. Via numerical experiments on transportation systems in California, we demonstrate the accuracy, computational efficiency, and robustness of the proposed approach compared to the Monte Carlo approach.

Motivation & Objective

• To enable rapid, scalable, and accurate seismic reliability assessment of highway bridge networks at the node level, beyond system-level metrics.
• To overcome the high computational cost of Monte Carlo simulations in large-scale infrastructure reliability analysis.
• To develop a surrogate model that supports inductive learning for unseen graphs, such as expanded or damaged networks.
• To quantify connectivity probabilities between all origin-destination pairs under probabilistic earthquake scenarios and bridge conditions.
• To demonstrate the model’s robustness and generalization capability on real-world transportation networks in California.

Proposed method

• The model uses a graph neural network (GNN) with message-passing mechanisms to learn node-level connectivity from graph-structured bridge network data.
• Node features include bridge-specific attributes (e.g., age, material), while edge features represent failure probabilities under seismic events.
• The GNN is trained end-to-end to predict binary connectivity outcomes (connected or disconnected) between all origin-destination pairs.
• The model leverages inductive bias through message passing, enabling prediction on unseen graphs without retraining.
• Training data is generated via Monte Carlo simulations to provide ground-truth connectivity probabilities for supervision.
• The framework supports transfer learning and generalization, tested by applying a Level 2-trained model to predict connectivity in a larger, unseen Level 3 region.

Experimental results

Research questions

• RQ1 Can a GNN surrogate model accurately predict node-level connectivity in highway bridge networks under seismic hazards, outperforming traditional Monte Carlo simulations in speed and scalability?
• RQ2 How well does the GNN model generalize to unseen graphs, such as those with newly added or removed bridges, without fine-tuning?
• RQ3 What is the performance of the GNN model when edge features are perturbed, simulating real-world uncertainty in bridge failure probabilities?
• RQ4 How does the model’s predictive accuracy vary with different levels of training data coverage (e.g., 60% vs. 100% of nodes)?
• RQ5 Can the GNN framework effectively support rapid reliability assessment for emergency planning and infrastructure resilience under extreme events?

Key findings

• The GNN surrogate model achieves high accuracy with F1 scores of 0.958, 0.949, and 0.961 across three regions when tested on perturbed edge features, indicating strong robustness to input uncertainty.
• The correlation coefficient between GNN predictions and Monte Carlo simulations reaches 0.942, 0.923, and 0.969 across the three regions, confirming strong predictive consistency.
• The model maintains an F1 score of 0.92 in inductive learning tasks, successfully predicting connectivity in a larger, unseen Level 3 region using a model trained only on a smaller Level 2 region.
• The GNN framework reduces computational time significantly compared to Monte Carlo simulations by eliminating the need for extensive sampling.
• The model demonstrates effective generalization to out-of-distribution data, including graphs with modified edge features and unseen topologies.
• The message-passing mechanism in the GNN enables modular, scalable learning that supports inductive transfer, making it suitable for dynamic infrastructure scenarios.