1 Introduction

Geotechnical engineering involves investigating and utilizing naturally occurring materials, including soil, rock, and intermediate geomaterials, such as coal [1, 2]. Among these materials, soil is distinguished due to its complex physical, mechanical, and chemical properties in engineering materials [3, 4]. These materials exhibit inherent anisotropic and heterogeneous characteristics resulting from various origins and formation mechanisms, presenting difficulties in understanding and forecasting [5, 6]. Traditionally, geotechnical engineers employ two primary approaches for investigating material behaviors: (1) laboratory and field tests and (2) numerical and analytical methods [7, 8]. While laboratory and field tests offer descriptive insights, they often entail substantial costs and time commitments [7, 9]. Conversely, numerical methods, like finite elements [10,11,12] or discrete analyses [12, 13], provide cost-effective virtual assessments of geotechnical material behavior [7, 14].

Computational intelligence and soft computing analyses have gained recognition due to the complex challenges encountered in various engineering applications. These approaches have gradually replaced the need for complex calculations [15,16,17]. There are numerous advantages to employing AI techniques in geotechnical engineering [18], including:

  1. 1.

    AI can model intricate and nonlinear processes without presuming initial input–output relationships [19, 20].

  2. 2.

    AI demonstrates its effectiveness in forecasting, surveillance, choice-making, recognition, and classification in various situations [21].

  3. 3.

    AI has the capability to provide precise predictions even when there are no established physical parameter relationships available [22].

  4. 4.

    AI has the ability to process extensive datasets, identify patterns, and occasionally generate missing data [23, 24].

Artificial neural networks (ANNs) can evaluate all feasible alternatives for a given project outcome using complex mathematical models and advanced software tools [25, 26]. The integration of ANNs with optimization algorithms is essential to mitigate error rates, particularly in complex scenarios like compressed sensing [27, 28]. ANN provides essential tools for geotechnical engineers in prominent consulting firms, enabling them to make quick and informed decisions, thereby improving performance and mitigating risks [29].

Geotechnical challenges are full of uncertainties and include different factors that avoid direct determination by engineers, leading to the quick adoption of machine learning (ML) techniques [30,31,32]. ML techniques can recognize potential correlations in data without any prior presumptions [33,34,35,36]. Additionally, deep learning (DL), a subfield of ML, aims to enhance the learning algorithms' capability to comprehend complex data. This is achieved using ANNs with multiple layers of interconnected nodes [37]. While DL has exhibited success in tackling learning challenges, its performance is influenced by various factors, and optimizing DL remains an ongoing focus of research in the field of AI [38, 39]. Furthermore, as computational efficiency advances, ongoing investigations into AI and DL are taking place [40, 41].

The primary objectives of this research are to comprehensively assess the applications of ANN, ML, DL, and EL in geotechnology forecasting and to establish a systematic categorization framework. Through the analysis of an extensive dataset, this study aims to provide insights into utilizing these techniques in addressing geotechnical challenges, enabling informed decision-making in this field. Table 1, which serves as an abbreviation table, provides crucial references to assist readers in understanding the fundamental ideas presented in the paper. Figure 1 illustrates various sections covered in this review paper.

Table 1 Explanations of Abbreviations Employed throughout the Paper
Full size table
Fig. 1
figure 1

Outline of various sections covered in the current review paper

Full size image

2 Literature review

Geotechnical engineering is a multidisciplinary field that encompasses various sub-disciplines within engineering and geology [1, 8]. It involves the study of soil and rock behavior to ensure the stability, safety, and longevity of infrastructure and construction projects [8,9,10]. In structural engineering, it addresses foundation design and soil–structure interaction [42, 43]. Construction engineering involves ground structures, excavation, soil improvement, and earthwork [8, 44]. Environmental engineering focuses on geoenvironmental concerns, while earthquake engineering deals with seismic geotechnics and ground motions [8, 9, 45].

Mechanical engineering aspects include rock mechanics, soil mechanics, and ice mechanics [9, 46]. Geology plays a role in geological engineering, geomaterials analysis, and geohazard assessment [8, 9]. Hydraulic engineering covers earth dams, scouring, groundwater drainage, and marine geotechnics [8, 47], while transportation engineering includes tunneling and road engineering [8, 9]. Figure 2 illustrates these diverse sub-disciplines within the field of geotechnical engineering. Geotechnical engineers apply their expertise across these domains, ensuring the proper utilization of soil and rock properties in diverse construction and environmental contexts.

Fig. 2
figure 2

Overview of geotechnical engineering sub-disciplines

Full size image

AI consists of a sophisticated collection of programming techniques [48, 49]. Many of these techniques are founded on the idea that knowledge gaining, organization, access, and modification, in both humans and machines, form the basis for 'intelligent' decision-making [50,51,52,53,54]. AI techniques find application in a wide array of geographical issues, including modeling individual and collective decision-making and developing expert and 'intelligent' geographical information systems [55]. Geotechnical engineers employ various AI techniques to solve diverse challenges [56]. Adopting AI applications in geotechnical engineering has revolutionized the resources available to industry experts, providing them with advanced tools for in-depth data analysis and intricate modeling [57, 58] decision-making [59]. Recent instances of GeoAI endeavors involve the identification of terrain features [60, 61], the detection of densely distributed building footprints [62,63,64], the extraction of information from scanned historical maps [65,66,67], and semantic classification, such as with LiDAR point clouds [68,69,70], novel methods for spatial interpolation [71], and advances in traffic forecasting [72,73,74]. Integrating AI applications in this field enhances the analytical capabilities of industry professionals and fundamentally alters their decision-making processes [75,76,77,78,79]. Through precise data analysis and the application of dynamic modeling, AI enables professionals to optimize site selection, fine-tune design specifications, and adeptly anticipate and manage risks, ultimately leading to the successful and sustainable execution of geotechnical projects [80]. AI plays a pivotal role in advancing sustainable construction and infrastructure projects by efficiently allocating resources, reducing environmental impacts, and optimizing material usage, energy consumption, and waste management techniques in geotechnical engineering [81,82,83]. It is a powerful tool in sustainable construction, effectively managing resources to minimize environmental impacts [84,85,86,87]. By optimizing material distribution and utilization, reducing energy consumption, and limiting waste, AI not only results in cost savings but also significantly diminishes the ecological footprint in construction [88,89,90]. Furthermore, AI can also be utilized to enhance other critical factors, such as mechanical strength and bearing capacity [91, 92]. This comprehensive approach utilizes AI’s capabilities to address a broader spectrum of considerations in construction, resulting in improved sustainability and performance. Nevertheless, the integration of AI in geotechnical engineering faces challenges, mainly due to the necessity for comprehensive and reliable data, particularly in specialized or remote projects [18, 93,94,95]. Ensuring data quality is essential, emphasizing the significance of a balanced approach in developing and validating AI models [96,97,98]. Table 2 discusses some review articles by researchers using AI in geotechnical engineering from 2017 onwards. Also, these reviews have concentrated on one or two AI techniques. In contrast, this review article offers a comprehensive exploration of all four AI techniques (ANN, DL, ML, and EL) within the field of geotechnical engineering. The study showed a co-occurrence keyword analysis encompassing AI techniques (ANN, DL, ML, and EL), systematic review, geotechnical engineering, and review; the data were gathered from the Scopus database and then visualized utilizing VOS Viewer. The dimensions and annotations of each circle represent the importance of the corresponding keyword. Lines connecting them represent connections between these keywords. Various colors signify separate clusters, each associated with its own specialized domain of knowledge. Figure 3 visually represents the research trend observed from 2020 to 2023.

Table 2 Review articles using AI techniques in geotechnical engineering
Full size table
Fig. 3
figure 3

Application of AI methods in review papers within the field of geotechnical engineering, supported by total publications retrieved from the Scopus database

Full size image

3 AI techniques and algorithms overview

The advent of big data, cloud computing, artificial neural networks, and machine learning has empowered engineers to develop machines capable of emulating human intelligence [107, 108]. Expanding on these advancements, this research designates machines capable of perceiving, recognizing, learning, reacting, and problem-solving as AI [109,110,111]. This inevitably signifies a transformative influence on future workplaces, as AI has the potential to enhance human performance to higher standards [112,113,114]. Consequently, it is poised to emerge as the next groundbreaking innovation [115]. AI is classified into four distinct approaches, including artificial neural networks (ANN), machine learning (ML), deep learning (DL), and ensemble learning (EL). The categorization of these methods is depicted in Fig. 4. Furthermore, for clarity, Table 3 offers a comprehensive comparison of these techniques across various dimensions. As demonstrated in Table 3, the assessment of complexity, data requirements, and interpretability can vary depending on the specific architecture and algorithm. These characteristics can be influenced by factors such as data quality and domain expertise.

Fig. 4
figure 4

Various categorizations of AI

Full size image
Table 3 Comparing the advantages and disadvantages of ANN, ML, DL, and EL
Full size table

3.1 Artificial neural network (ANN) and its application in geotechnical engineering field

The development of the ANN appeared as a solution for tackling challenges involving complex patterns and predictions [120, 121]. Inspired by the information processing mechanisms of the human brain, studies have defined the complex, multi-layered structure of ANN [122, 123]. These neural networks consist of three essential layers: the input, hidden, and output [124]. Neurons are distributed across these layers in a multilayer ANN, each neuron serving as a crucial processing unit. The initial level, represented by the input layer, acquires information to reduce errors and enhance computations [125, 126]. Consequently, the logical determination of the number of neurons is crucial. The input signal can move to subsequent layers due to the interconnectivity among neurons. Neuron weight signifies their capacity to communicate with one another; moreover, the weight and neuron count in preceding layers determine the number of neurons in each layer [127, 128]. It is worth noting that the discretion of the number of hidden layers and neurons is possible. Like other networks, ANNs serve as an exceptional modeling tool for analysis. They excel in defining nonlinear network function evaluation, pattern recognition, data classification, simulation, clustering, and optimization, all essential features of AI [129]. ANN can be categorized into six distinct network types, which include:

  • Feed Forward Neural Network (FFNN) FFNN is a foundational framework within supervised ANNs, demonstrating notable proficiency in recognizing patterns [130]. It systematically handles information through input, hidden, and output layers linearly, devoid of feedback connections [131]. It is skilled at complex pattern learning, although it requires precise adjustment of hyperparameters to achieve the best results.

  • Back Propagation Neural Network (BPNN) The Backpropagation Neural Network (BPNN) is a widespread ANN used in supervised learning tasks. It demonstrates proficiency in comprehending complex relationships [132]. Functioning through interlinked layers, it refines weights using backpropagation to reduce output differences [133].

  • Radial Basis Function Neural Network (RBFNN) The RBFNN is a neural network model designed for various ANN applications [134,135,136]. It utilizes radial basis functions. This architecture enables it to excel in both pattern recognition and regression tasks [137]. By employing radial basis functions, it efficiently processes data and adjusts parameters dynamically, ensuring accurate and reliable results [138, 139].

  • Bayesian Regression Neural Network (BRNN) The BRNN combines neural networks with Bayesian regression to represent complex models [140, 141]. It utilizes neural networks to manage nonlinear patterns and employs Bayesian methods to measure uncertainty, making it advantageous for various applications [142].

  • Generalized Regression Neural Network (GRNN) The GRNN is a sophisticated model proficient in making predictions by estimating functions through a radial basis function strategy [143, 144]. This feature makes it especially apt for efficient training and approximating smooth functions [145,146,147]. The GRNN, based on in radial basis functions (RBFs), is recognized for its effectiveness in regression assignments.

  • Differentiated Evolution Neural Network (DENN) DENN is a type of ANN that uses the differential evolution algorithm to optimize its network structure and parameters. The DENN integrates advanced evolution strategies in neural network training to accelerate convergence speed, improve solution quality, and enhance generalization capabilities for complex optimization tasks [148, 149].

ANN can be employed for a range of tasks in geoengineering, including Soil Classification [93, 150, 151] and Property Estimation [93, 152, 153], Settlement and Settlement Prediction [93, 154,155,156], Slope Stability Analysis [157,158,159], Seismic Hazard Assessment [160, 161], Groundwater Flow Modeling [162, 163], Tunneling and Excavation [164,165,166], Site Characterization [95, 167], Risk Assessment [168, 169], Material Behavior Modeling [170, 171], and Optimization [172, 173].

Different types of ANNs, including GRNN, DENN, BRNN, RBFNN, and FFNN, may be chosen based on the specific problem, data availability, and the desired level of complexity. For example, RBFNNs may be used for data interpolation and function approximation, while FFNNs are suitable for general regression and classification tasks. DENN, if applicable to geotechnical problems, may offer specific advantages in terms of optimization and adaptation [174].

Table 4 displays the employment of ANN techniques in geotechnical engineering.

Table 4 Using ANNs Techniques in Geotechnical Engineering
Full size table

From the information provided, it is clear that a diverse array of advanced AI techniques, including various types of neural networks and hybrid models, have been successfully utilized in research within the field of geotechnical engineering [104, 192]. These approaches have addressed various geotechnical challenges, from soil property prediction to estimating material strengths and evaluating geotechnical structure performance. The outcomes substantiate the efficacy of AI-based models in providing accurate and dependable forecasts across different facets of geotechnical engineering. Furthermore, these models offer the potential to improve computational efficiency and make valuable contributions to advancing more sustainable practices in soil stabilization and subgrade construction. [8, 100, 104, 176].

The study performed a keyword analysis, giving particular attention to the application of ANN techniques in the field of Geotechnical Engineering. The data were gathered from the Scopus database and then visualized utilizing VOS Viewer. Over the period from 2016 to 2023, a total of 1254 manuscripts were cumulatively published. The size and label of each circle correspond to the significance of the respective keyword. Connecting lines indicate relationships between the keywords. Different colors denote distinct clusters based on their specific areas of expertise, which is presented in Fig. 5. Furthermore, based on data from the WOS database, a geographic analysis demonstrates the utilization of ANN techniques in geotechnical engineering between 2016 and 2023, as depicted in Fig. 6.

Fig. 5
figure 5

Keywords related to ANN in the field of geotechnical engineering, extracted from the Scopus database

Full size image
Fig. 6
figure 6

Utilization of ANN techniques in the analysis of geotechnical engineering, evidenced by total publications categorized by country in the WOS database

Full size image

3.2 Machine learning (ML) and its application in geotechnical engineering field

ML represents a vital advancement in AI [121]. ML is achieved through iterative algorithms that learn from relevant data specific to a particular training task. This enables computers to recognize complex patterns and bring to light insights without the need for direct programming [193]. ML aims to automate analytical modeling, especially for tasks involving high-dimensional data, such as classification, regression, and clustering [121]. Different varieties of ML models contain:

Reinforcement learning: reinforcement learning involves training an agent to interact with its environment using feedback signals, aiming to develop a strategy that maximizes anticipated rewards. As indicated by [194,195,196,197,198,199,200,201,202,203,204,205,206,207,208], this type of ML can be classified into four methods.

  1. 1.

    Value-based methods focus on acquiring value functions (e.g., Q-values or state values) to guide action selection based on these estimates.

  2. 2.

    Policy-based methods directly learn policies to choose actions that lead to maximum expected rewards.

  3. 3.

    Actor-critic methods combine estimating value functions with policy optimization.

  4. 4.

    Model-based methods entail learning a model of the environment to plan and make decisions.

Unsupervised learning: unsupervised learning involves identifying patterns or structures in data without prior knowledge of the desired outcome. It is trained on data that lack labels, aiming to learn a representation that captures the inherent structure of the dataset [194,195,196,197,198]. This learning includes a variety of techniques, such as clustering, dimensionality reduction, density estimation, and anomaly detection [209,210,211]. Clustering groups similar data points based on specific features or similarities [196, 212, 213]. Dimensionality reduction methods aim to reduce the number of features while retaining important information [214,215,216,217]. Density estimation focuses on estimating the probability density function of a dataset [218,219,220]. Anomaly detection identifies data points that deviate significantly from expected or normal behavior [210, 221, 222]. Some well-known and commonly used techniques in Unsupervised Learning, such as Principal Component Analysis (PCA) [223,224,225], K-Means Clustering [226,227,228], Hierarchical Clustering [229, 230], Gaussian Mixture Models (GMM) [231,232,233], are mentioned here.

Supervised learning: supervised learning trains an algorithm using labeled data, where each input corresponds to a known output. The algorithm learns to link input features with desired outcomes through these labeled examples. This learning process is divided into two main types: classification, which categorizes data into predefined classes, and regression, which predicts continuous numerical values. Classification yields distinct class labels, while regression deals with a range of continuous outputs [194,195,196,197,198,199, 234, 235]. Some widely recognized and commonly used techniques in supervised learning are mentioned here, including Linear Regression [236], Logistic Regression (LR) [237, 238], Bayesian Linear regression (BLR) [239, 240], Random Forest [241, 242], Support Vector Machines (SVM) [243, 244].

ML techniques, including supervised, unsupervised, and reinforcement learning, have a wide range of applications in geoengineering [245, 246]. These techniques can enhance decision-making, optimize processes, and gain geotechnical and geographical data insights [247]. Supervised learning can be applied in slope stability analysis, foundation design, and material classification [248,249,250], while clustering for site characterization and dimensionality reduction uses unsupervised learning [251, 252]. Additionally, reinforcement learning is applied for optimal excavation tunneling and resource management [253, 254]. The successful application of ML in geoengineering depends on the availability of high-quality data, domain expertise, and careful model selection and validation. Table 5 is dedicated to ML methodologies designed to tackle specific challenges in geotechnical engineering.

Table 5 Application of Machine Learning Methodologies in Geotechnical Challenges
Full size table

Table 4 offers a summary of recent studies exploring the utilization of ML in geotechnical engineering. These investigations contain various subjects, ranging from soil classification and spatial interpolation to slope stability and rock mass categorization. Also it contains predictions for unconfined compressive strength (UCS), evaluations of soil layering, projections for shear strength of fiber-reinforced soil (FRS), estimations of cation exchange capacity (CEC), and assessments of gully erosion susceptibility. The results of these inquiries demonstrate the effectiveness of ML algorithms in dealing with various challenges within geotechnical engineering. ML models have demonstrated notable accuracy in tasks like soil classification, spatial property variability prediction, slope stability assessment, rock mass categorization, UCS prediction, identification of soil layers, FRS shear strength prediction, CEC estimation, and gully erosion susceptibility mapping [106, 192, 269]. Furthermore, the research emphasizes the significance of factors such as the quality and representativeness of the training dataset, model complexity, and the specific application context when deploying ML algorithms in geotechnical engineering. Additionally, further validation of ML models using new databases is often necessary to evaluate their broader applicability. According to a search of the Scopus database for Elsevier journal papers, researchers published 1,401 research papers on ML in geotechnical engineering between 2016 and 2023. Figure 7 shows these data, along with keywords related to Machine learning techniques in geotechnical engineering, extracted from the most relevant articles. Moreover, as indicated by the WOS database, a geographical data analysis demonstrates the application of ML techniques in geotechnical engineering between 2016 and 2023, as shown in Fig. 8.

Fig. 7
figure 7

ML keywords in geotechnical engineering from the Scopus database

Full size image
Fig. 8
figure 8

Utilization of ML techniques in the analysis of geotechnical engineering, evidenced by total publications categorized by country in the WOS database

Full size image

3.3 Deep learning (DL) and its application in geotechnical engineering field

Recently, DL has demonstrated remarkable advancements and achievements across a wide range of fields [270]. DL, a subset of ML, aims to develop algorithms that can gradually comprehend complex data representations. This is accomplished by employing neural networks consisting of interconnected layers of nodes. DL algorithms typically use extensive datasets during training, enabling them to identify complex patterns and attain highly accurate predictions [271]. DL algorithms possess the unique capability to autonomously identify features, circumventing the need for ML algorithms, which accelerates data classification processes [121]. Additionally, DL demonstrates exceptional efficiency in handling substantial volumes of information within tight timeframes. Notably, one of the most noteworthy characteristics of DL is its capacity to enhance its intelligence over time continually [270]. Table 5 provides an overview of DL methods. Other DL approaches often integrate and complement the methods outlined in Table 6 to enhance overall efficiency and effectiveness. Also, Table 7 is specifically dedicated to the application of DL techniques in the field of geotechnical engineering.

Table 6 Comprehensive Overview of DL Methods
Full size table
Table 7 Application of Deep Learning Methodologies in Geotechnical Engineering
Full size table

DL techniques have found applications in various aspects of geoengineering due to their ability to process and analyze complex data patterns, including geological feature detection [93, 105, 286], landslide prediction[169, 287, 288], seismic data analysis [105, 289], groundwater modeling [290], infrastructure monitoring [93, 291], soil classification[292, 293], geospatial data analysis [105, 290], mining and resource management [294, 295], environmental impact assessment [105, 296]. To implement DL in geoengineering, access to relevant datasets, machine learning and deep learning expertise, and computing resources for model training will be required [297, 298].

The research conducted a keyword analysis with a specific emphasis on the utilization of Deep Learning techniques in Geotechnical Engineering. It was found that researchers published 1,040 research papers on deep learning in this field between 2016 and 2023. The data were collected from the Scopus database and visualized using VOS Viewer, as illustrated in Fig. 9; this graphical representation captures the evolving research trends spanning from 2019 to 2023. Moreover, as indicated by the WOS database, a geographical data analysis demonstrates the application of DL techniques in geotechnical engineering between 2016 and 2023, as shown in Fig. 10.

Fig. 9
figure 9

DL keywords in geotechnical engineering from the Scopus database

Full size image
Fig. 10
figure 10

Utilization of DL techniques in the analysis of geotechnical engineering, evidenced by total publications categorized by country in the WOS database

Full size image

3.4 Ensemble learning (EL) its application in geotechnical engineering field

EL, a method in ML, combines the forecasts of multiple models to enhance overall performance [299, 300]. EL aims to improve predictive performance, accuracy, and generalization on various tasks [301, 302]. Ensemble methods work best when the base models are diverse, meaning they make errors on different subsets of the data or have different approaches to solving the problem [303, 304]. This diversity helps in reducing the overall error [305, 306]. Several widely recognized EL techniques include:

Bagging (Bootstrap Aggregating): This ensemble method involves training multiple models on different subsets of the data, and their predictions are aggregated [307,308,309]. Random Forest [310, 311] and Bagged Decision Trees [312,313,314] are the well-known methods in this category.

Boosting: Boosting enhances predictive performance by training weak models sequentially. Each model corrects the mistakes made by its predecessor, resulting in a strong learner within the ensemble [315, 316]. AdaBoost (Adaptive Boosting) [317,318,319], Gradient Boosting Machines (GBM) [320,321,322], XGBoost (Extreme Gradient Boosting) [323,324,325], LightGBM (Light Gradient Boosting Machine) [326, 327], and CatBoost (Categorical Boosting) [328, 329] are widely recognized techniques within this classification.

Stacking Ensembles (SE): In this approach, a meta-model is trained to learn how to best combine predictions from the base models [330, 331]. Stacking Classifier [332,333,334] and Stacking Regressor [335, 336] are recognized techniques within the SE category.

Voting Ensembles (VE): Models in this ensemble provide predictions, and a majority vote determines the final output [337, 338]. Hard Voting [339,340,341] and Soft Voting [342,343,344] are established methods within the VE classification.

In geotechnical engineering, EL technique is frequently employed to heighten the precision of soil and rock behavior predictions [345]. In geotechnical engineering, stacking is a frequently used EL method [346, 347]. This involves training multiple ML models on the same dataset and combining their predictions to generate a conclusive forecast, which can be executed through techniques like weighted averaging or voting [348]. Another common technique utilized in geotechnical engineering is Bagging [349,350,351]. Multiple ML approaches are trained on distinct subsets of the dataset, and their predictions are aggregated to form a final prediction. This helps mitigate the overfitting of the models to the training data [352,353,354].

It is important to note that the choice of EL method and the specific application will depend on the nature of the geoengineering problem, the available data, and the goals of the analysis [93, 286, 355, 356]. EL can significantly enhance the predictive capabilities and robustness of models in geoengineering, ultimately leading to safer and more effective engineering solutions [357,358,359]. EL is a developing field ready to have a revolutionary impact on geotechnical engineering. Table 8 provides a collection of recent studies that have successfully utilized EL to address a range of challenges in the field of geotechnical engineering. These include forecasting soil liquefaction susceptibility, categorizing rock mass quality, approximating lateral wall deflection in braced excavations, projecting soil properties through raw soil spectra data, and anticipating landslide susceptibility. These efforts emphasize the potential of EL in increasing the accuracy, efficiency, and reliability of geotechnical analyses and designs [357, 363].

Table 8 Using of Ensemble Learning Technique in Geotechnical Engineering
Full size table

The data, sourced from the Scopus database, were subsequently visualized using VOS Viewer. Over the period from 2016 to 2023, researchers published 609 research papers on ensemble learning in geotechnical engineering. The size and label of each circle in the visualization indicate the significance of the respective keyword, while connecting lines signify relationships between them. Figure 11 presents these data along with keywords associated with Ensemble Learning (EL) approaches in geotechnical engineering, extracted from the most pertinent articles. Furthermore, according to the WOS database, the application of EL techniques in geotechnical engineering is demonstrated through geographical data analysis, as depicted in Fig. 12, which visually depicts the research pattern observed from 2020 to 2023.

Fig. 11
figure 11

Keywords related to EL in the field of geotechnical engineering, extracted from the Scopus database

Full size image
Fig. 12
figure 12

Utilization of DL techniques in the analysis of geotechnical engineering, evidenced by total publications categorized by country in the WOS database

Full size image

To gain insight into the performance of various EL-based models in the geotechnical field, Kardani et al. [293] examined the effectiveness of different EL techniques in predicting the resilient modulus of subgrade soils. They found that the bagging ensemble model outperformed other models tested, including the voting ensemble, voting ensemble with random forest, and stacking ensemble. Their conclusion was that the bagging ensemble outperformed other methods, making it suitable for estimating the resilient modulus with superior performance and an acceptable degree of accuracy. This model not only demonstrated higher prediction accuracy and generalization ability, but also exhibited several advantages such as stability, reduced noise, and ease of use. On the other hand, learning the art of ensemble modeling can be challenging, and making incorrect selections may lead to reduced prediction precision. Additionally, ensemble modeling can be costly in terms of both time and space. However, additional research using various datasets should be conducted to predict different geotechnical parameters, ensuring the performance of the bagging ensemble methods and other EL-based methods. Therefore, it is strongly recommended to utilize EL-based methods in the geotechnical field for predicting mechanical, physical, and chemical properties of soils. Further research is necessary to make reliable decisions about their performance in the geotechnical area.

4 Discussions and challenges linked with AI in geotechnical engineering

ANN models are adaptable and capable of capturing complex patterns in data. However, these models require precise adjustment of hyperparameters to achieve peak performance [116]. The performance of ANN depends on factors such as architecture, data quality, and data quantity [124, 365, 366]. Therefore, ANN is able to be the best choice for small datasets or when interpretability is crucial.

Various techniques are included in ML models, such as supervised learning, unsupervised learning, or reinforcement learning, which are well suited for different tasks. The performance of ML models varies according to the algorithm and the data being utilized. In comparison to DL models, ML models are frequently found to be more interpretable [367, 368]. They are considered a favorable option when there are limited data or the need for transparent models [117].

DL models, such as CNNs and RNNs, manage extensive, high-dimensional datasets proficiently. They can automatically acquire hierarchical features [369, 370]. DL demands significant computational power, sizable datasets, and precise parameter optimization. Additionally, DL models may not always offer interpretability, which can present limitations in certain use cases [118].

EL combines multiple models to enhance predictive performance, often surpassing the performance of individual models. It achieves this by reducing overfitting and increasing robustness, making it suitable for diverse datasets and applications [119]. Furthermore, EL demonstrates a reduced susceptibility to noise and outliers [371].

Assessing ANN, ML, DL, and EL for geoengineering in terms of accuracy and performance can be a complex task, given that the efficacy of each method relies on diverse variables, such as the particular problem, dataset characteristics, and model setup. These approaches rely on geoengineering problems, the data, computational resources, and interpretability needs. Both ML and ANN demonstrate a moderate level of complexity and are mainly applied in the field of geotechnics. Notably, ML has attracted substantial attention from researchers due to its high interpretability and optimal performance, even with small data. This interest is substantiated by data from WOS covering the period from 2019 to 2023, which reveals that a significant number of articles published in Springer Nature, Elsevier, and IEEE journals within the geotechnical domain underscore the prevalent preference for employing ML among researchers in this field as shown in Fig. 13.

Fig. 13
figure 13

Comparative analysis of esteemed journals (Springer, Elsevier, and IEEE) in the domains of ANN, ML, DL, and EL in geoengineering, 2019–2023, using the WOS database

Full size image

Figure 13 illustrates the number of research papers published in reputable journals, such as those from Springer, Elsevier, and IEEE, focusing on the areas of ANN, ML, DL, and in the field of Geotechnical Engineering. These data have been sourced from WOS.

Based on data from the WOS database, ANN is frequently employed in geotechnical engineering, even though ML, DL, and EL methods have demonstrated substantial potential as illustrated in Fig. 14. This preference for using ANN in geotechnical engineering may be attributed to the common requirement for real-world laboratory data frequently encountered by civil and geotechnical engineers or the potential limitation in expertise for effectively employing ML, DL, and EL methods in data-driven prediction. However, EL techniques consistently outperform the other three methods in the context of predicting geotechnical behaviors.

Fig. 14
figure 14

Comparison of various ANN, ML, DL, and EL methods used in geotechnical engineering based on the total number of publications from 2019 to 2023, using the WOS database

Full size image

According to the data obtained from the WOS database, Fig. 14 provides an overview of the utilization of ANN, ML, DL, and EL approaches in geotechnical engineering from 2019 to 2023. The data clearly show that ANN has maintained its status as a consistently preferred technique within this field. Additionally, it is noteworthy that ML has exhibited a steady and upward trend over the years. In 2022 and 2023, researchers demonstrated a nearly equal preference for both ANN and ML techniques within the field of geotechnical engineering.

As depicted in Fig. 14, it is evident that the EL methods have been consistently popular over the years. Notably, the utilization of EL in the field of geotechnical engineering experienced a substantial increase from 2021 to 2022, reaching its peak adoption rate during this period. DL methods have not been widely adopted in recent years, but they started gaining recognition in geotechnical engineering in 2020. However, their popularity among geotechnical engineers remains limited due to the substantial amount of data required for accurate forecasting using this learning approach.

A widespread trend toward the utilization of artificial intelligence techniques, including ANN, ML, DL, and EL, in the field of geotechnical engineering is observed globally. This analysis, spanning from 2016 to 2023, involves the classification of data using WOS enabling thorough examination of transformations on a continental scale (refer to Fig. 15). This comparison reveals that this subject matter is actively embraced across all continents.

Fig. 15
figure 15

Continental comparison of ANN, ML, DL, and EL in geotechnical engineering (2016–2023) using WOS database

Full size image

5 Future research directions and opportunities linked with AI application in geotechnical engineering

Future research in the field of geotechnical engineering and artificial intelligence (AI) should prioritize interdisciplinary collaboration, bringing together geotechnical engineering expertise and AI proficiency. This synergy has the potential to yield innovative solutions and provide a deeper understanding of how AI can effectively address the multifaceted challenges within geotechnical engineering. Furthermore, researchers should investigate geographical variations in the utilization of AI techniques in geotechnical engineering, examining how these methods are applied differently in various regions and identifying the factors influencing these variations. Additionally, the integration of AI for real-time monitoring and decision-making during geotechnical construction and operations should be explored, focusing on the development of adaptive AI-driven systems that can enhance safety and operational efficiency. Finally, researchers should delve into the concept of human–machine collaboration, examining how AI can assist geotechnical practitioners in decision-making, risk assessment, and project design. These research directions, aligned with the standards of scholarly articles, aim to foster innovation and provide practical solutions for the geotechnical engineering community. Figure 16 offers a visual depiction of the critical future research direction in the application of AI within the realm of geotechnical engineering.

Fig. 16
figure 16

Visual overview of key future research directions in the application of AI within the realm of geotechnical engineering

Full size image

From a geotechnical engineering perspective, there are numerous topics that can still be studied and addressed in future research. One potential area of research is the application of various AI methods to predict the dynamic response of different soils, contingent on the availability of adequate datasets. In addition, a simple review of Tables 4, 5, 7, and 8 and available papers in the field of geotechnical engineering confirms that soil improvement, as a hot topic in general, has received less attention from the AI approach. It is well known that soil properties, including soil gradation, consistency, compaction parameters, consolidation, dispersivity, collapsibility, swelling potential, durability, strength, elasticity, stress–strain curves, peck strain energy, resilient modulus, dynamic response, erodibility, chemical compositions, hydraulic conductivity, electrical conductivity, and liquefaction potential, can be altered through stabilization with traditional materials like lime and cement, or through the use of waste by-products such as lignosulfonate, travertine waste, red mud, sewage sludge, water treatment sludge, fly ash, various types of slags, as well as soil reinforcement using different materials like fibers and geosynthetic materials, or alternative soil improvement techniques such as electroosmosis [46, 372,373,374,375,376,377,378,379,380,381,382,383,384,385]. However, it is evident that AI-based prediction of soil parameters after stabilization or reinforcement with various techniques and materials deserves more attention, especially considering the substantial number of experimental papers in this field and the availability of sufficient datasets. Therefore, future research studies can focus on the prediction of stabilized and reinforced soil parameters.

6 Conclusions

ANN, ML, DL, and EL are pivotal approaches for extracting valuable insights and making autonomous predictions in various fields, including geotechnology. This study aimed to comprehensively assess the applications of these techniques in geoengineering, filling a critical gap in the existing literature.

Evaluation of a vast dataset extracted from the Web of Science and Scopus databases revealed significant insights. ANN remains a widely used technique in geotechnical engineering, often due to the necessity for real-world laboratory data frequently encountered by civil and geotechnical engineers. Additionally, the expertise gap in effectively applying ML, DL, and EL methods for data-driven predictions may influence the preference for ANN. However, when it comes to predicting geotechnical behaviors, EL techniques consistently outperform the other three methods, showcasing their effectiveness in this domain.

Each of these techniques possesses its unique strengths and limitations. ANN models are adaptable and excel at capturing complex data patterns, but they require meticulous hyperparameter tuning and are suitable for scenarios with limited data or where interpretability is crucial. ML models encompass various techniques suitable for diverse tasks, offering interpretable solutions and being favored when data are limited. DL models handle high-dimensional data effectively but demand substantial computational resources and careful parameter optimization. Conversely, EL combines multiple models to enhance predictive performance, exhibiting robustness and reduced sensitivity to noise and outliers. The integration of ANN, ML, DL, and EL techniques has significantly contributed to advancing the field of geotechnology. Researchers and practitioners in this domain should continue to explore and harness the potential of these methodologies to address the evolving challenges in geotechnical engineering effectively.