Enhancing Bridge Pier Scour Depth Prediction Through Bi-Directional LSTM Networks: A Physics-Informed ML Learning Approach

Permutation importance analysis reveals that flow depth (y) is the dominant predictor (R² drop = 0.55-0.70),consistent with its role in governing near-bed velocity and vortex strength around piers.Permutation feature importance for BiLSTM model CONCLUSIONBidirectional LSTM networks establish themselves as highly effective tools for physics-informed temporal scour prediction, substantially outperforming conventional recurrent architectures and traditional empirical approaches. Key advantages over traditional methods:Real-time predictive capability without extensive computational overhead Superior generalization across diverse hydraulic conditions Computational efficiency compared to resource-intensive CFD simulations Captures complex nonlinear scour dynamics effectively. Future Research DirectionsFuture work should integrate attention mechanisms to enhance temporal interpretability and extend the framework to complex pier configurations including pile groups and bridge confluences.Deployment in real-time scour monitoring systems with adaptive sensor data assimilation will enable proactive bridge safety management and early warning capabilities.