Application of Integrated Machine Learning and Digital Twins for Real-Time Monitoring and Slope Stability Analysis

This study developed an integrated machine learning and digital twin system to monitor and predict slope stability in real-time using Random Forest (RF), Support Vector Machine (SVM), Extreme Gradient Boosting (XGBoost), and Long Short-Term Memory (LSTM) algorithms trained and tested on 847 field monitoring cases obtained during 18 months of an instrumented slope region in Southwest China. The dataset consisted of input parameters derived from displacement sensors, piezometers, and meteorological devices. For factor of safety (FoS) prediction, SVM achieved the highest accuracy in terms of coefficient of determination (R 2 = 0.8244) and root mean square error (RMSE = 0.0630), followed by RF (R 2 = 0.8048) and XGBoost (R 2 = 0.7576). The LSTM network failed in regression (R 2 = − 0.3675), producing near-constant predictions that did not capture temporal variations in FoS. For stability classification, XGBoost attained the highest F1-score (0.955) and accuracy (96.09%), while SVM detected all marginally stable cases with 100 percent minority-class recall. However, LSTM's apparent classification accuracy (95.19%) was entirely attributable to majority-class bias, with zero recall for the marginally stable minority class, a result fully consistent with its regression failure. Feature importance analysis showed that shallow and deep pore water pressures collectively accounted for 71–75 percent of predictive importance. A digital twin platform was developed enabling integration of real-time sensor data, physics-based models, and three-dimensional visualization for processing monitoring data. The results indicate that conventional ML models (SVM, RF, XGBoost) provide reliable FoS predictions for operational slope monitoring, whereas the LSTM architecture as configured is unsuitable for both regression and safety–critical classification tasks.