Journal of Science and Transport Technology

Enhancing landslide susceptibility prediction using deep learning through optimal activation and loss function selection

Tran Huu Anh — 2026-10-05

Landslide susceptibility prediction has increasingly shifted from traditional statistical methods toward deep learning approaches capable of modeling complex nonlinear relationships among geo-environmental controlling factors. However, model performance is influenced not only by network architecture but also by the selection of activation and loss functions, which directly affect learning dynamics and generalization capability. This study systematically investigates the influence of activation–loss function combinations within a unified deep learning framework for landslide susceptibility mapping in Sin Ho area, northwestern Vietnam. A multi-source geospatial database integrating topographic, geological, hydrological, land-use, seismic, and rainfall-related factors was used to train and validate multiple model configurations. Model performance was evaluated using a comprehensive set of quantitative indicators, including area under the ROC curve (AUC) and accuracy (ACC). The results indicate that models employing the Softmax activation function consistently achieve superior predictive performance and robustness. Among all tested configurations, the Softmax–MultiLabel (SM–ML) model delivered the most reliable results, attaining a validation AUC of 0.878, and ACC of 83.75, demonstrating strong discrimination ability, stable generalization, and reduced prediction error. The landslide susceptibility map generated using the optimal SM–ML configuration exhibits a clear and progressive increase in landslide density from very low to very high susceptibility classes, confirming meaningful spatial stratification of landslide hazard. These findings highlight that careful selection of activation and loss functions can substantially enhance deep learning performance without increasing model complexity or computational cost, providing a robust and efficient framework for landslide susceptibility assessment and disaster-risk reduction in mountainous regions under changing climatic conditions.

An efficient hybrid framework for reliability-based size and shape optimization of trusses considering uncertainties

Anh Hoang Nguyen — 2026-10-04

Reliability-based design optimization (RBDO) of truss structures has become a powerful tool for achieving optimal designs under uncertainties in material properties, cross-sectional areas, and applied loadings. By integrating the sequential optimization and reliability assessment (SORA) into an efficient optimization tool based on the success-history-based adaptive differential evolution (SHADE) algorithm, this paper presents a novel RBDO methodology, which is the so-called SORA-SHADE, for the optimization of shape and size in truss structures under uncertainties of input data. The proposed paradigm is tested on well-known 2D and 3D trusses with and without the reliability analyses for validation. For reliability-based analysis cases, each cross-sectional area is considered as a random variable under uncertainties. The results show that the proposed approach is highly effective and reliable in finding optimal designs that meet high reliability requirements.

A hybrid linear regression and CatBoost model for high-performance nonlinear static analysis of functionally graded plates

Do Thi Thanh Dieu — 2026-10-02

Nonlinear static analysis of functionally graded material (FGM) plates using high-fidelity numerical methods such as Isogeometric Analysis (IGA) is computationally demanding because it requires iterative solution procedures to trace the equilibrium path. Recently, deep learning surrogates (e.g., Long Short-Term Memory, LSTM) have been explored to reduce repeated simulations; however, they can still require nontrivial training effort and careful model tuning. This study proposes a high-performance hybrid surrogate model combining Linear Regression (LR) and CatBoost to predict the nonlinear load-deflection behavior of FGM plates. The equilibrium-path response is formulated as a sequence regression task along pseudo-time load steps and decomposed into two stages: LR captures the global linear trend, while CatBoost learns the nonlinear residual component not represented by LR. The proposed approach is trained and validated using IGA-generated datasets. It is evaluated under different data-availability settings (30%, 50%, and 80% of the equilibrium-path sequence used for training and testing) and assessed in terms of prediction accuracy, hyperparameter sensitivity, and end-to-end computational cost (training + prediction). In our implementation, the combined LR-CatBoost training and prediction time is approximately 0.25 s for the baseline case, excluding the IGA computation required to generate the initially available response segment. Prediction accuracy is high, with mean absolute percentage errors on the order of 10^-2 to 10^-1% depending on the case. These findings indicate that the proposed LR-CatBoost strategy provides a robust and ultra-fast surrogate for near-real-time nonlinear structural analysis of FGM plates.

Experimental study on the performance of a slotted wells turbine in an oscillating water column device

Thuc Tri Dang — 2026-10-14

Ocean wave energy has garnered significant attention as a sustainable power source, with the Oscillating Water Column (OWC) device emerging as a highly effective technology for wave-to-electricity conversion. Within this system, the Wells turbine serves as a critical component for energy extraction. While current literature predominantly relies on Computational Fluid Dynamics (CFD) simulations to analyze the performance of these turbines, this study employs a rigorous experimental approach to evaluate their dynamics. The experimental results demonstrate high reliability when compared with CFD simulation data. Furthermore, this research evaluates the performance of a slotted Wells turbine design. The comparative analysis reveals that the slotted Wells turbine achieves superior efficiency compared to the conventional Wells turbine design. These findings offer a comprehensive empirical foundation and provide critical insights for the future optimization of wave energy converters. Ultimately, this study serves as an essential precursor to subsequent developments aimed at enhancing the operational viability of OWC systems in practical, real-world applications.