Despite the emergence of recent advancements, machine learning (ML) based methods for estimating the fragility curves of structures through probabilistic ground motion selection techniques pose a challenge due to the computational cost associated with data preparation. The primary aim of this research is to reduce the data preparation time involved in estimating the fragility curves of structures using a ground motion selection approach that considers earthquake magnitude, distance from the seismic source, and shear wave velocity of soil as essential parameters. To achieve this objective, ML algorithms are employed to calculate the fragility curves of various reinforced concrete moment resisting (RC/MR) frames with different periods, utilizing codebased and generalized conditional intensity measure (GCIM) ground motion selection methods. The SMOTE-ENN technique, a data resampling method, is used to balance the training data for the ML algorithms to address potential bias resulting from imbalanced training data. To validate the fragility curves obtained through ML, analytical fragility curves are derived for a specific structure at three damage levels and compared with the ML curves. The results demonstrate that the percentage of the enclosed area between the analytical and ML curves, relative to the area under the analytical curve, is below 10 % and 5 % for the GCIM and code-based methods, respectively. Fragility curves were generated for various structures, including regular and irregular buildings, to investigate the generalizability. Results indicate that, for the specific structures analyzed in this study, excluding torsional ones, the structure's period is a sufficient structural feature for generating fragility curves.

Different slope geohazards have different causal mechanisms. This study aims to propose a method to investigate the decision-making mechanisms for the susceptibility of different slope geohazards. The study includes a geospatial dataset consisting of 1203 historical slope geohazard units, including slope creeps, shallow slides, rockfalls and debris flows, and 584 non-geohazard units, and 22 initial condition factors. Following a 7:3 ratio, the data were randomly divided into a test set and a training set, and an ensemble SMOTE-RF-SHAP model was constructed. The performance and generalization ability of the model were evaluated by confusion matrix and the receiver operating characteristic (ROC) for the four types of geohazards. The decision-making mechanism of different geohazards was then identified and investigated using the Shapley additive explanations (SHAP) model. The results show that the hybrid optimization improves the overall accuracy of the model from 0.486 to 0.831, with significant improvements in the prediction accuracy for all four types of slope geohazards, as well as reductions in misclassification and omission rates. Furthermore, this study reveals that the main influencing factors and spatiotemporal distribution of different slope geohazards exhibit high similarity, while the impacts of individual factors and different factor values on different slope geohazards demonstrate significant differences. For example, prolonged continuous rainfall can erode rock masses and lead to slope creep, increased rainfall may trigger shallow mountain landslides, and sudden surface runoff can even cause debris flows. These findings have important practical implications for slope geohazards risk management. (c) 2025 Institute of Rock and Soil Mechanics, Chinese Academy of Sciences. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/ 4.0/).