This paper presents a detailed review of existing and emerging deep learning algorithms for time series forecasting in geotechnics and geoscience applications. Deep learning has shown promising results in addressing complex prediction problems involving large datasets and multiple interacting variables without requiring extensive feature extraction. This study provides an in-depth description of prominent deep learning methods, including recurrent neural networks (RNNs), convolutional neural networks (CNNs), generative adversarial network, deep belief network, reinforcement learning, attention and transformer algorithms as well as hybrid networks using a combination of these architectures. In addition, this paper summarizes the applications of these models in various fields, including mining and tunnelling, railway and road construction, seismology, slope stability, earth retaining and stabilizing structures, remote sensing, as well as scour and erosion. This review reveals that RNN-based models, particularly Long Short-Term Memory networks, are the most commonly used models for time series forecasting. The advantages of deep learning models over traditional machine learning, including their superior ability to handle complex patterns and process large-scale data more effectively, are discussed. Furthermore, in time series forecasting within the fields of geotechnics and geosciences, studies frequently reveal that deep learning methods tend to surpass traditional machine learning techniques in effectiveness.

The soil water retention curve (SWRC) or soil–water characteristic curve (SWCC) is a fundamental feature of unsaturated soil that simply shows the relationship between soil suction and water content (in terms of the degree of saturation and volumetric or gravimetric water content). In this study, the applications of the SWRC or SWCC have been extensively reviewed, taking about 403 previously published research studies into consideration. This was achieved on the basis of classification-based problems and application-based problems, which solve the widest array of geotechnical engineering problems relevant to and correlating with SWRC geo-structural behavior. At the end of the exercises, the SWRC geo-structural problem-solving scope, as covered in the theoretical framework, showed that soil type, soil parameter, measuring test, predictive technique, slope stability, bearing capacity, settlement, and seepage-based problems have been efficiently solved by proffering constitutive and artificial intelligence solutions to earthwork infrastructure; and identified matric suction as the most influential parameter. Finally, a summary of these research findings and key challenges and opportunities for future tentative research topics is proposed.

This study focuses on identifying damage in reinforced concrete (RC) beams using time-domain modal testing and wavelet analysis. A numerical model of an RC beam was used to generate various damage scenarios with different severities and locations. Acceleration time histories were recorded for both damaged and undamaged structures. Two damage indices, DI_MW and DI_SW, derived from the wavelet analysis, were employed to determine the location and severity of the damage. The results showed that different wavelet families and specific mother wavelets had varying effectiveness in detecting damage. The Daubechies wavelet family (db2, db6, and db9) detected damage at the center and sides of the RC beams due to good time and frequency localization. The Biorthogonal wavelet family (bior2.8 and bior3.1) provided improved time–frequency resolution. The Symlets wavelet family (sym2 and sym7) offered a balanced trade-off between time and frequency localization. The Shannon wavelet family (shan1-0.5 and shan1-0.1) exhibited good time localization, while the Frequency B-Spline wavelet family (fbsp2-1-0.1) excelled in frequency localization. Certain combinations of mother wavelets, such as shan1-0.5 with the DI_SW index, were highly effective in detecting damage. The DI_SW index outperformed DI_MW across different numerical models. Selecting appropriate wavelet analysis techniques, particularly utilizing shan1-0.5 in the DI_SW, proved effective for detecting damage in RC beams.