structural failure, necessitating robust prediction and reliability assessment. This study develops a novel hybrid framework integrating Harris Hawks Optimization–Support Vector Regression (HHO-SVR) and a joint Long Short-Term Memory–Convolutional Neural Network (LSTM-CNN) with Monte Carlo simulation (MCS) for probabilistic reliability analysis. Using a comprehensive dataset of 153 laboratory clear-water scour observations, five dimensionless parameters—V/Vc, y/b, b/d50, Fr, and σg—were utilized to predict normalized scour depth (ds/b). The HHO algorithm effectively optimized SVR hyperparameters (C = 67.25, ε = 0.047, and γ = 0.326). While the HHO-SVR model achieved an R² = 0.543 and RMSE = 0.213, the LSTM-CNN model demonstrated superior generalization capability with an R² = 0.757 and RMSE = 0.156. Robustness was further verified through 5-fold cross-validation, yielding mean R² values of 0.578 and 0.738 for HHO-SVR and LSTM-CNN, respectively. MCS with 100,000 iterations revealed that the LSTM-CNN model attained a reliability index of β = 3.45$ at ds/b = 1.8, exceeding the structural safety threshold (β = 3.0), whereas HHO-SVR reached β = 2.46 at ds/b = 2.0. Sensitivity analysis identified V/Vc as the dominant factor (~67–70%), followed by b/d50 (~16–23%). The proposed framework is specifically validated for single circular piers in non-cohesive, relatively uniform sediments under clear-water conditions within defined parametric ranges (V/Vc: 0.40–0.99; b/d50: 18–280). These findings provide a high-fidelity tool for engineers to assess pier safety beyond deterministic predictions, although application to complex pier geometries or live-bed conditions requires further recalibration.