The macroscopic mechanical properties of granular systems largely depend on the complex mechanical responses of force chains at the mesoscopic level. This study offers an alternative to rapidly identify and predict force chain distributions under different stress states. 100 sets of gradation curves that effectively represent four typical continuous gradation distributions are constructed. Numerical specimens corresponding to these gradation curves are generated using the discrete element method (DEM), and a dataset for deep neural network training is established via biaxial compression numerical simulations. The relationship between particle distribution characteristics and force chain structure is captured by the Pix2Pix conditional generative adversarial network (cGAN). The effectiveness of the generated force chain images in reproducing both particle gradation and spatial distribution characteristics is verified through the extraction and analysis of pixel probability distributions across different color channels, along with the computation of texture feature metrics. In addition, a GoogLeNet-based prediction model is constructed to demonstrate the accuracy with which the generated force chain images characterize the macroscopic mechanical properties of granular assemblies. The results indicate that the Pix2Pix network effectively predicts and identifies force chain distributions at peak stress for different gradation

Landslides are destructive natural disasters that cause catastrophic damage and loss of life worldwide. Accurately predicting landslide displacement enables effective early warning and risk management. However, the limited availability of on-site measurement data has been a substantial obstacle in developing data-driven models, such as state-of-the-art machine learning (ML) models. To address these challenges, this study proposes a data augmentation framework that uses generative adversarial networks (GANs), a recent advance in generative artificial intelligence (AI), to improve the accuracy of landslide displacement prediction. The framework provides effective data augmentation to enhance limited datasets. A recurrent GAN model, RGAN-LS, is proposed, specifically designed to generate realistic synthetic multivariate time series that mimics the characteristics of real landslide on-site measurement data. A customized moment-matching loss is incorporated in addition to the adversarial loss in GAN during the training of RGAN-LS to capture the temporal dynamics and correlations in real time series data. Then, the synthetic data generated by RGAN-LS is used to enhance the training of long short-term memory (LSTM) networks and particle swarm optimization-support vector machine (PSO-SVM) models for landslide displacement prediction tasks. Results on two landslides in the Three Gorges Reservoir (TGR) region show a significant improvement in LSTM model prediction performance when trained on augmented data. For instance, in the case of the Baishuihe landslide, the average root mean square error (RMSE) increases by 16.11%, and the mean absolute error (MAE) by 17.59%. More importantly, the model's responsiveness during mutational stages is enhanced for early warning purposes. However, the results have shown that the static PSO-SVM model only sees marginal gains compared to recurrent models such as LSTM. Further analysis indicates that an optimal synthetic-to-real data ratio (50% on the illustration cases) maximizes the improvements. This also demonstrates the robustness and effectiveness of supplementing training data for dynamic models to obtain better results. By using the powerful generative AI approach, RGAN-LS can generate high-fidelity synthetic landslide data. This is critical for improving the performance of advanced ML models in predicting landslide displacement, particularly when there are limited training data. Additionally, this approach has the potential to expand the use of generative AI in geohazard risk management and other research areas. (c) 2024 Institute of Rock and Soil Mechanics, Chinese Academy of Sciences. Production and hosting 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/).

This study focused on exploring the utilization of a one-part geopolymer (OPG) as a sustainable alternative binder to ordinary Portland cement (OPC) in soil stabilization, offering significant environmental advantages. The unconfined compressive strength (UCS) was the key index for evaluating the efficacy of OPG in soil stabilization, traditionally demanding substantial resources in terms of cost and time. In this research, four distinct deep learning (DL) models (Artificial Neural Network [ANN], Backpropagation Neural Network [BPNN], Convolutional Neural Network [CNN], and Long Short-Term Memory [LSTM]) were employed to predict the UCS of OPG-stabilized soft clay, providing a more efficient and precise methodology. Among these models, CNN exhibited the highest performance (MAE = 0.022, R2 = 0.9938), followed by LSTM (MAE = 0.0274, R2 = 0.9924) and BPNN (MAE = 0.0272, R2 = 0.9921). The Wasserstein Generative Adversarial Network (WGAN) was further utilized to generate additional synthetic samples for expanding the training dataset. The incorporation of the synthetic samples generated by WGAN models into the training set for the DL models led to improved performance. When the number of synthetic samples achieved 200, the WGAN-CNN model provided the most accurate results, with an R2 value of 0.9978 and MAE value of 0.9978. Furthermore, to assess the reliability of the DL models and gain insights into the influence of input variables on the predicted outcomes, interpretable Machine Learning techniques, including a sensitivity analysis, Shapley Additive Explanation (SHAP), and 1D Partial Dependence Plot (PDP) were employed for analyzing and interpreting the CNN and WGAN-CNN models. This research illuminates new aspects of the application of DL models with training on real and synthetic data in evaluating the strength properties of the OPG-stabilized soil, contributing to saving time and cost.

Droughts are among the most devastating natural hazards, occurring in all regions with different climate conditions. The impacts of droughts result in significant damages annually around the world. While drought is generally described as a slow-developing hazardous event, a rapidly developing type of drought, the so-called flash drought has been revealed by recent studies. The rapid onset and strong intensity of flash droughts require accurate real-time monitoring. Addressing this issue, a Generative Adversarial Network (GAN) is developed in this study to monitor flash droughts over the Contiguous United States (CONUS). GAN contains two models: (a) discriminator and (b) generator. The developed architecture in this study employs a Markovian discriminator, which emphasizes the spatial dependencies, with a modified U-Net generator, tuned for optimal performance. To determine the best loss function for the generator, four different networks are developed with different loss functions, including Mean Absolute Error (MAE), adversarial loss, a combination of adversarial loss with Mean Square Error (MSE), and a combination of adversarial loss with MAE. Utilizing daily datasets collected from NLDAS-2 and Standardized Soil Moisture Index (SSI) maps, the network is trained for real-time daily SSI monitoring. Comparative assessments reveal the proposed GAN's superior ability to replicate SSI values over U-Net and Na & iuml;ve models. Evaluation metrics further underscore that the developed GAN successfully identifies both fine- and coarse-scale spatial drought patterns and abrupt changes in the SSI temporal patterns that is important for flash drought identification. A new deep learning-based model using a generative adversarial network (GAN) is developed for real-time flash drought detection and monitoring Remote sensing maps are used as inputs to encompass the entire regions within the CONUS The proposed GAN is able to capture abrupt changes in drought patterns