Bridge Superstructure Deterioration Prediction with Uncertainty Quantification through Modern Hopfield Network–Based Conformal Predictors

Transportation agencies require accurate predictions of bridge superstructures to effectively manage bridge assets. Conventional deterioration models for predicting bridge superstructure conditions typically produce single-label classifications without assessing the confidence or uncertainty associated with each prediction step. This limitation impairs the ability to distinguish between high-confidence predictions, which are suitable for immediate action, and low-confidence predictions that require further inspection. This study introduces a conformal prediction (CP) approach using modern Hopfield networks (MHNs) to quantify the uncertainty in bridge superstructure condition prediction. The proposed framework was applied to predict future bridge superstructures, and bidirectional long short-term memory (Bi-LSTM), when combined with conformal prediction, achieved 85% accuracy while maintaining a narrow prediction set (mean size of 1.12). This finding suggests that the conditions of most bridge superstructures can be predicted with high confidence. The CP framework enables transportation agencies to make data-driven maintenance, repair, and rehabilitation decisions by quantifying the uncertainty in the predicted bridge superstructures.