The integration of Artificial Intelligence (AI) in geotechnical engineering has revolutionized the study of reactive soils, which are known for their significant volume changes due to moisture variations. This article presents a scientometric analysis that traces the evolution of AI applications in reactive soil research over the past three decades, revealing a steady increase in the adoption of AI techniques. Key AI methods such as artificial neural networks (ANN), support vector machines (SVM), fuzzy logic, genetic algorithms (GA), and image analysis have been instrumental in soil characterisation, strength prediction, and stabilisation. The analysis identifies emerging trends and opportunities, including the potential for real-time geo-structural health monitoring, multi-disciplinary forecasting, and the development of sustainable geo-structural designs. By addressing current gaps and leveraging advanced AI techniques, this study paves the way for more accurate predictive models and innovative solutions to mitigate the impacts of reactive soils on infrastructure. The article also highlights the importance of improving algorithm transparency and interpretability to make AI-driven solutions more practical and reliable for real-world applications.
AI Generated
This summary of the content was generated with the help of AI.
Abstract
Reactive soils present significant challenges in geotechnical engineering due to their unpredictable behaviour, which can severely impact buildings and infrastructure. While artificial intelligence (AI) has been applied to improve numerical modelling of reactive soils, the scope of its application and future potential remains unclear. AI methods, such as neural networks, support vector machines, genetic algorithms, fuzzy logic, and image analysis, have shown promise in soil characterisation, strength prediction, performance evaluation, clay cracking analysis, and soil stabilisation. However, a systematic understanding of these advancements is lacking. This research addresses the gap by conducting a scientometric analysis using tools like VOSviewer®, Citespace®, and Sci2® to map scientific knowledge, identify trends, and uncover future opportunities. Findings suggest that integrating nanotechnology, real-time monitoring, multidisciplinary forecasting, and shared knowledge databases can enhance AI applications. This analysis provides a foundation for advancing AI-driven solutions in geotechnical engineering and addressing the challenges posed by reactive soils.
Notes
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
1 Introduction
1.1 Evolution of AI in Geotechnical Engineering
Over the last 35 years, AI has been increasingly adopted in geotechnical engineering, driven by advancements in mimicking natural searching and optimisation algorithms. AI refers to the concept of developing computer systems that are capable of tasks commonly requiring human intelligence (Stevenson 2010). AI tasks may include recognising objects, understanding speech, responding to a conversation, solving problems, optimising solutions, greeting people, and driving a vehicle (Theodoridis 2020). An application of AI is collectively called machine (ML), the process of implementing mathematical models to help a computer or robot learn or perform a specific task by improving measured performance without direct instruction or hard coding (Samuel 1959; Mitchell 1998).
Between 1980 and 2000, Geotechnical investigations that used the theory of AI and ML algorithms were limited, but the rise of cutting-edge technologies in data availability, storage, and computing steadily increased the application of AI from 2001 to 2020. An average of seven papers were annually published from 1980 to 2000 and now almost seventy papers are produced in the area for the last two decades per year, with the highest record in 2020 equal to 184 articles. Early AI techniques applied in geotechnical engineering were expert systems, fuzzy logic, and pattern recognition for assessing abutments, landslides, building foundations, and mines (Gupta and Bodechtel 1982; Scoble et al. 1986; Adams et al. 1989; Wong et al. 1989; Katsuumi et al. 2024). The application of AI extended to the use of an artificial neural network (ANN), hybrid expert systems, and image analysis of soil parameters and geo-structures (Maher and Williams 1991; Chan et al. 1995; Kayen et al. 1999; Oliphant 1999). The rise of cutting-edge technologies in data availability, storage, and computing has steadily increased the application of AI to hazard mitigation, geo-structural health monitoring, and nanotechnology using Deep Learning (DL) and Convolutional Neural Network (CNN), clustering, hybrid genetic algorithms, and fuzzy logic (Kadivar et al. 2011; Abbaszadeh Shahri 2016b; Amezquita-Sanchez et al. 2016; Congress and Puppala 2020).
Advertisement
1.2 Clayey Soils and Their Challenges
Compared to other types of soil being studied in the geotechnical field, clay has observed significant interest for researchers since it commonly exhibits a lower mechanical strength, particularly in wetter states and greater volumetric variations due to changes in moisture content (Ural 2018). Clayey soils are ubiquitous and can be found in many areas around the globe, particularly in America (the United States of America, Canada, Mexico, Guatemala, Venezuela, Colombia, Peru, Bolivia, Venezuela, Brazil, Argentina), Europe (the United Kingdom, Spain), Africa (Egypt, Sudan, Ethiopia), Asia (Turkey, India, China, Malaysia) and Australia.
Most clayey soils are reactive, which undergo significant volume changes due to variations in soil water content, leading to ground swelling when their water content increases and shrinking when their water content decreases (Teodosio 2020). Such swelling and shrinking behaviour induce damage to lightweight structures such as pavements, underground pipes, and residential structures (Johnson 1969; Petry and Little 2002; Teodosio et al. 2020a). Reactive soils-induced distress to physical infrastructures has been reported globally, including in Australia, China, Egypt, India, Israel, South Africa, the United Kingdom, and the United States of America, resulting in significant socio-economic impact (Li et al. 2014). Infrastructure rehabilitation and construction expenditures as a result of soil movements are more than twice the loss incurred from natural disasters such as floods, hurricanes, tornadoes, and earthquakes (Jones and Jefferson 2012). Addressing these challenges requires innovative solutions, highlighting the growing role of AI in mitigating reactive soil impacts.
1.3 Knowledge Gaps and Exploring Potential Field Advancements
The complex challenges brought by reactive soils require multi-physical non-linear analysis for a heterogeneous and anisotropic material. The application of AI algorithms to reactive soils can allow the processing of “big data”, build adaptive non-linear models, and predict more insightful outcomes whilst considering randomness (Theodoridis 2020). The first recorded AI application to reactive soil was conducted by Hallaire (1993), a decade after the first AI application in geotechnical engineering, according to the Scopus database. Applications of AI in reactive soil research were related to soil characterisation and strength prediction, soil and structure performance, clay cracking and desiccation, and soil movement and stabilisation (Gong et al. 2004; Mahfouz et al. 2007; Shengquan et al. 2015; Yin et al. 2018; Huang et al. 2019). Regularly occurring AI techniques utilised in reactive soil research were artificial neural networks, support vector machines, genetic algorithms, fuzzy logic, and image analysis (Das et al. 2010; Samui et al. 2011; Mozumder and Laskar 2015; Julina and Thyagaraj 2019; Congress and Puppala 2020).
Despite these advancements, the application of AI in reactive soil research remains limited, and a comprehensive understanding of its development and potential is lacking. This research seeks to address this gap through a scientometric analysis, which evaluates the evolution of AI applications in reactive soil studies and identifies emerging trends and opportunities. By mapping past and current research directions, this study provides a foundation for future advancements in the field and highlights innovative solutions for managing the challenges posed by reactive soils.
Advertisement
2 Methodology
A scientometric analysis is a type of science mapping through bibliometric tools, data, and methods aiming to detect the intellectual structure of a specific research domain (Darko et al. 2020). This scientometric analysis used geospatial analysis, term co-occurrence, bibliographic coupling, and citation bursts using VOSviewer® (van Eck and Waltman 2018), Citespace® (Chen 2017), and Sci2® (Linnemeier et al. 2010) to identify research on the application of AI to study reactive soils. This analysis method is useful for presenting a visual representation of significant trends in a large volume of literature datasets (Chen 2017). The methodology of this scientometric analysis is comprised of the following stages: bibliometric tool selection, record collection, and data analysis (Bonitz 1994).
2.1 Scientometric Analysis
The scientometric analysis in this research was conducted using four main stages illustrated in Fig. 1. These stages give the context of the current research and future direction in applying AI to reactive soil. The first stage presents the location where most studies were conducted using geospatial analysis. The geospatial analysis in this study used geolocation attributes of the collected papers from their affiliation sections. If a publication has multiple authors with different affiliations, the details of the first author are considered. This analysis gives a context of where publications related to the application of AI to reactive soil challenges were conducted and will determine the countries or regions actively researching the said topic.
Fig. 1
Four stages of the scientometric analysis used in this study
The second stage shows the main research interest in the specialised field using topical analysis. The topical analysis using the term co-occurrence evaluates the incidence and frequency of two unique terms used in the title, abstract, and keywords sections (Linnemeier et al. 2010). This technique identifies the topic similarity of a basic and aggregate research domain related to AI and reactive soils. Publications that have shared more terms are considered to have high topical overlap, which is connected using links and placed closer to each other. The minimum weight for the considered terms was set to fifteen.
The third stage demonstrates the citation patterns through network analysis. The network analysis gives an understanding of the natural and man-made complex interrelations of publications. There are three main ways to derive networks from the bibliometric data collected; these are the direct citation network, the co-citation clustering, and bibliographic coupling (Kleminski et al. 2020). Direct citation network connects publications that cite each other, as shown in Fig. 2a, where publications A, C, and D and publications A, B, and D are connected. The second way, co-citation clustering, connects publications cited by the same publications as presented in Fig. 2b. In this example, publications C and D are connected since A cited both items. In contrast, bibliographic coupling connects publications that are cited by the same publication. For instance, in Fig. 2c, publication D was cited by both A and B, thus making these papers connected. This scientometric study used bibliographic coupling since this technique is most applicable to short-term analyses and the scientific domain (Klavans and Boyack 2017).
Fig. 2
Three ways to derive bibliometric networks: a direct citation network, b co-citation clustering, and c bibliographic coupling
Lastly, the fourth stage determines the trending topics over time, employing citation burst analyses. Citation burst analysis is an indicator of the most active research area in a specific date range. This technique detects publication events associated with a surge of citations due to trending topics (Zhou et al. 2019). This study implemented burst analyses for both terms and citation in the publication to determine what topics and which journals have been considered in the growing interest of applying AI to reactive soils.
2.2 Bibliometric Tool Selection
Various bibliometric tools can be selected to perform a scientometric analysis. However, each tool has its advantages and disadvantages (Cobo et al. 2011). There are three general aspects to consider in performing a scientometric analysis, being the capability of data pre-processing, the ability of network analysis, and the visualisation options (Moral Muñoz et al. 2020). To perform a reliable scientometric survey, VOSviewer® (van Eck and Waltman 2018), Citespace® (Chen 2017), and Sci2® (Linnemeier et al. 2010) were selected due to their functionality of implementing varying clustering techniques. VOSviewer®, developed by van Eck and Waltman (2014), is software used to construct and view bibliometric maps. The advantage of VOSviewer® is its intuitive pre-processing interface and intelligible outputs. However, this convenience limits the analysis to general clustering and analysis that can be further grouped to acquire more insightful results (Darko et al. 2020). Citespace®, developed by Chen (2006), is a bibliometric software focusing on more detailed network analysis and modifiable visuals. Citespace® can analyse and present temporal clustering that may not be available in other bibliometric tools. Sci2®, developed by Börner et al. (2003), is a powerful tool for massive bibliometric data pre-processing, network analysis, and visual layout customisation. VOSviewer®, Citespace®, and Sci2® were used to perform different scientometric analyses on the citation records, enabling the identification of current research and future direction in applying AI to reactive soil research. Each tool has unique strengths and weaknesses, as summarised below:
VOSviewer®: Offers intuitive preprocessing and clear visual outputs but is limited to general clustering analyses (van Eck and Waltman 2014).
Citespace®: Specializes in temporal clustering and modifiable visualisations (Chen 2006).
Sci2®: Capable of extensive data preprocessing, network analysis, and customisable layouts (Börner et al. 2003).
These tools were selected for their complementary features and were used for varying aspects of clustering and visualisation to ensure a robust analysis of reactive soil research.
3 Data Collection
Scholarly databases are increasingly important in the academic ecosystem (Guerrero-Bote et al. 2021; Liu et al. 2024). For this research, citation records were collected from the Scopus database due to the availability and coverage of peer-reviewed scientific publications when compared to Web of Science or Google (Darko et al. 2020). Furthermore, Scopus has a relatively expedited indexing process that is beneficial when considering novel studies such as topics related to AI. Meho and Rogers (2008) concluded that Scopus can be used as an exclusive database source for scientometric analysis due to its consideration of a wide range of publication types and applicability to the scientific research domain. For this reason, the study used Scopus as the citation data source to review the development of AI methods applied to reactive soils. Several terms were chosen to identify the publications that were related to both AI and reactive soils. The selection of the terms was based on the common terminologies used by researchers in the research area. These terms were detected in the title, abstract, and keyword sections of each publication. The terms used related to AI were “machine learning”, “artificial intelligence”, “neural network”, “deep learning”, “computational intelligence”, “pattern recognition”, “convolutional neural network”, “recurrent neural network, “statistical learning”, “adaptive signal processing”, “image analysis”, “speech recognition”, “machine intelligence”, “expert systems”, “case-based reasoning”, “data mining”, “fuzzy logic”, “fuzzy sets”, “robotics”, “knowledge-based systems”, “support vector machine”, and “artificial general intelligence”. These keywords were adapted from Theodoridis (2020) and Salehi and Burgueño (2018). The terms related to reactive soils that were used with the AI terms are “expansive soil”, “reactive soil”, “shrink-swell soil”, “clay soil”, and “clayey soil”. There were no filters specified for the date range and the article type to cover a wider period and relevant peer-refereed sources (Meho and Rogers 2008). To compare the application of AI to reactive soil and the entire geotechnical field, another collection was obtained using the terms related to AI, described above, and the term “geotechnical engineering”.
A total of 254 peer-reviewed publications were collected for research related to the application of AI to reactive soils, whilst 1,918 papers were for the application of AI to geotechnical engineering. Each collected publication was evaluated based on the title, abstract, and keyword to check the relevance to this scientometric study. The number of publications was reduced to 148 and 1663 after disregarding research not related to AI applied to reactive soils and geotechnical engineering, respectively. The omitted papers were commonly associated with geoenvironmental engineering, which is outside the scope of the analysis of this study.
The AI algorithms applied in the current literature is summarised in Table 1. The advantages and disadvantages are listed for each AI technique, as well as their common applications.
Table 1
Summary of different AI algorithms in the literature
Algorithm
Advantages
Disadvantages
Possible applications
Image analysis
Accuracy in object detection, recognition, and segmentation
Can automate analyses
Requires a substantial amount of annotated data for training
Computationally intensive
Interpretability of complex models can be challenging
Landslide detection, rock mass characterization, and feature recognition
Identifying potential foundation issues caused by soil expansion and shrinkage
Artificial neural network (ANN)
Effective for applications with clear class separation
Robust against overfitting
Handles high-dimensional datasets
Deep architectures may suffer from the vanishing gradient problem
Identify potential areas prone to soil expansion and shrinkage
Support vector machine (SVM)
Effective for applications with clear class separation
Robust against overfitting,
Handles high-dimensional datasets
May not perform well with overlapping classes/imbalanced datasets
Efficiency may decrease for large-scale datasets
May require additional post-processing
Soil type classification
Rock mass classification
Landslide hazard mapping
Identify reactive soil areas
Assess the potential risks
Genetic algorithms (GA)
Optimisation problems
Can find better solutions
Robust in handling noisy and complex datasets
Convergence speed can be slow
Computationally expensive
Efficiency dependent on the design of the fitness function and problem representation
May struggle with high-dimensional optimisation spaces
Optimising pile placement
Slope stability analysis
Parameter calibration
Optimise the design of shallow and deep foundations
Fuzzy logic
Handles uncertainty and imprecision in data
Useful for dealing with vague or incomplete information
Can be subjective, labour-intensive and computationally intensive
May not be suitable for applications requiring precise and deterministic outcomes
Soil classification with uncertain properties, risk assessment, and decision-making in foundation settlement and geotechnical hazard management
4 Results and Discussion
4.1 Past and Present Research Trends
Almost a decade after the first AI application in geotechnical engineering, Hallaire (1993) applied AI in reactive soil research, according to Scopus. Hallaire (1993) used image analysis for microcrack orientation in clayey soil. Figure 3 presented the trend in research publications on the application of AI in reactive soils, in which there was an average of three papers per year from 1993 to 1996; however, this paused until 2001. From 2004 to the present, the average annual publication on AI-reactive soils increased from eight to twenty-five papers. The number of published research is increasing with peaks occurring. The distribution of published research indicates an increased application of AI in reactive soil research. Noting that technological advancements, availability of data and increased computing power may all be attributable.
Fig. 3
The annual number of published articles on reactive soils compared with that on geotechnical engineering (including reactive soil investigations) applying AI techniques based on the Scopus database
The AI techniques and ML algorithms used in reactive soil research were image analysis, artificial neural networks, support vector machines, genetic algorithms, and fuzzy logic discussed in the following sections.
4.1.1 Image Analysis
Image analysis was introduced in the 1960s to automatically recognise patterns in images and videos (Fernández-Caballero et al. 2008). The image analysis recognises fundamental attributes such as shapes, edges, objects, and texture to extract insightful information (Dang et al. 2019). Tasks related to image analysis can be as non-complex as bar code identification or as advanced as face recognition (Wei et al. 2011). The first application of AI to reactive soil research in the Scopus database was conducted by Hallaire (1993) using image analysis for microcrack orientation in clayey soil. This topic paved the way for the further investigation of soil cracking and desiccation conducted by Costa et al. (2008), Nakano et al. (2013), Singh et al. (2018).
CNN, a specific type of image analysis combined with neural networks, has acquired research interest in various fields of study. CNN uses a convolution that slides a filter over the input layer to perform image detection techniques commonly implemented in facial recognition, picture classification, document digitisation, and other image-based applications (Shin et al. 2016). CNN has been implemented in different applications in geotechnical engineering, mainly in studying clay microstructure and geo-structural health monitoring of abutments, tunnels, and pavements (Abdel-Qader et al. 2003; Basma et al. 2003; Dorafshan et al. 2019; Koch et al. 2015). In spite of its application to geotechnical engineering, CNN has not been used for reactive soil research.
4.1.2 Artificial Neural Network
An ANN is a computing technique designed to emulate the neural network of a human brain in processing and analysing information (Guresen et al. 2011). A collection of connected nodes called artificial neurons comprise a network connected by links called edges to transmit signals from a single neuron to other neurons. These signals are represented by real numbers called weights that are optimised to obtain the lowest value of a cost function. Application of AI techniques to reactive soil research mostly used ANN algorithms for regression analyses.
The applications of ANN to reactive soil research have mainly focused on strength and behaviour prediction, property determination, and settlement and evaluation. In the early 2000s, Basma et al. (2003) researched on time-dependent swelling of clays using ANN and was followed by the studies of Doris et al. (2008), Ashayeri and Yasrebi (2009), Ikizler et al. (2010), Bekhor and Livneh (2014), and Dounane and Trouzine (2020) to predict the swell percentage of reactive soils. In the late 2000, the application of ANN in determining reactive soil properties was also observed. For instance, Ozer et al. (2008) used ANN to determine the compression index of clayey soils whilst Aladag et al. (2013) and Ziaie Moayed et al. (2018) have estimated parameters related to limit pressure and modulus. Other studies used ANN for clay property determination (Bahmed et al. 2019) and swell-strength prediction (Jalal et al. 2021). Most of the applications of ANN were related as well to soil stabilisation of reactive soils such as papers by Mozumder and Laskar (2015), Onyelowe et al. (2021), Ramakrishna et al. (2011), and Sabat (2015). Other ANN implementations were related to the geo-structural health monitoring model (Kaya et al. 2018), landslide susceptibility (Abbaszadeh Shahri 2016a), slope failure (Abdalla et al. 2015), and raft-pile foundation settlement (Liu et al. 2020).
4.1.3 Support Vector Machine
SVM is a supervised learning AI technique with the objective of finding a hyperplane in a dimensional place that distinctly classifies the data points with the ability to generalise datasets (Dibike et al. 2000). Due to this strength in generalisation, most SVM applications to reactive soil research were related to classification analyses. For example, Ma (2005) and Yang and Liu (2012) used SVM to classify various reactive soils. SVM was also used for the prediction of geo-structural stone columns of highway embankments (Aljanabi et al. 2018) and reactive soil stabilisation using concrete admixtures (Sabat 2015) and geopolymer (Mozumder et al. 2017).
4.1.4 Genetic Algorithms
A GA is a heuristic technique inspired by the theory of natural evolution and concepts of genetics to solve search and optimisation problems. (Oreta et al. 2012). It is a stochastic search algorithm having a population that is sampled simultaneously, imitating natural biological processes, such as inheritance, mutation, selection and crossover (Aguilar-Rivera et al. 2015). GA was applied to some reactive soil research, such as identifying soil grade (Shi and Guo 2009), calculating settlement and bearing capacity of soil and foundation (Ziaie Moayed et al. 2018), predicting load-settlement behaviour of raft-pile foundation (Liu et al. 2020), and estimating the swell strength of reactive soils (Jalal et al. 2021).
4.1.5 Fuzzy Logic
Fuzzy logic is a form of many-valued logic in which the truth value of variables may be any real number between 0 and 1, employing the concept of partial truth, as opposed to Boolean logic that can only be 0 or 1 (Liao 2003). It is used in natural language processing and modern control expert systems considering uncertainty to develop a system with fast response and low computational power (Wasantha et al. 2012). Fuzzy logic is commonly used with neural networks to mimic the decision-making of a human for an application requiring faster processing (Cheng et al. 2012).
An early application of fuzzy logic to reactive soil research was conducted by Gong et al. (2004) to assess the reliability of roadbed settlements constructed on reactive soils. Lu et al. (2006) and Jalal et al. (2021) utilised Adaptive Neuro-Fuzzy Inference Systems (ANFIS) to classify and predict the shrinking and swelling of reactive soils. Subsequently, Wang and Chen (2010) and Yin et al. (2018) developed a model using fuzzy logic to evaluate the shrinking and swelling of reactive soils. Recently, Ikeagwuani and Nwonu (2020) applied fuzzy logic, together with the grey-based Taguchi approach, to optimise the number of additives in treating reactive soil.
Image analysis, ANN, SVM, GA, and fuzzy logic were used in reactive soil research based on the data collected from the Scopus database. The following sections will discuss the four main stages of the scientometric analysis implemented in this research. These stages will give the context of the current research and future direction in applying AI to reactive soil. The first stage presents the location where most studies were conducted using geospatial analysis. The second stage shows the main research interest in the specialised field using topical analysis. The third stage demonstrates the citation patterns through network analysis. The fourth stage determines the trending topics over time, employing citation burst analyses.
4.2 Stage 1: Geospatial Analysis
The geospatial analysis gives the geographical locations of where studies were conducted. Geospatial analysis was performed using Sci2®. Geolocation attributes were collected based on the first authors’ affiliation details. Note that only the first author’s detail was used to reduce the complexity of the analysis. This spatial geolocation analysis also incorporated the temporal factor. Each country presented in Fig. 4 was grouped, and the average publication year was calculated. This analysis shows the proportion of studies for each country considering the activity in the considered period. Hence, lighter colour indicates that most research activities in a specific country are concentrated in early 2000, whilst darker colour indicates studies that are recently conducted in that particular country.
Fig. 4
Number of published papers by country using geospatial analysis of research on reactive soil employing AI
Most countries presented in Fig. 4 had recorded substantial research related to the shrinking and swelling movement of reactive soils (Li et al. 2014). These regions are in line with the clayey areas shown in Fig. 1 (NASA 2021). The areas with clayey soil in both Figs. 1 and 4 primarily reflect the need to further investigate reactive soil and apply AI to confront the challenges brought by its disruptive shrink-swell movement.
The majority of studies applying AI to reactive soil investigations were from China (30%), India (24%), Turkey (12%), Iran (11%), the US (10%), and Canada (6%), as shown in Fig. 4 and Table 2. These countries also have darker colours in Fig. 4, which indicate recent activities in studying the possible usage of AI to alleviate reactive soil impacts. In contrast, Germany, Spain, Egypt, France, the United Kingdom, and Sweden have lighter colours, indicating that the average number of research publications is around the early 2000s. This reflects that investigations related to AI and reactive soils have not been continuously conducted in most European countries (i.e., Germany, Spain, France, the United Kingdom, and Sweden). This may perhaps be due to the location of most European countries not having many challenges with reactive soil, with the exception of the United Kingdom (Jones et al. 2010).
Table 2
Publications grouped by country based on research affiliation
Gong et al. (2004), Miao and Yu (2004), Shi et al. (2004), Ma (2005), Lu et al. (2006), Mahfouz et al. (2007), Wen (2007), Tang et al. (2008), Ding et al. (2009), Wang and Chen (2010), Xia and Wu (2014), Shengquan et al. (2015), Dai et al. (2016), Yin et al. (2018), Huang et al. (2019), Yang et al. (2019), Zhao et al. (2019), Li et al. (2019)
(Viswanadham et al. (2009), Das et al. (2010), Rajesh and Viswanadham (2011), Ramakrishna et al. (2011), Samui et al. (2011), Mozumder and Laskar (2015), Muduli et al. (2015), Mondal et al. (2016), Vibhute et al. (2016), Anupam et al. (2017), Mozumder et al. (2017), Bhattacharya et al. (2018), Singh et al. (2018), Julina and Thyagaraj (2019), Borthakur and Dey (2020)
(Goktepe et al. (2008), Ozer et al. (2008), Ikizler et al. (2010, 2014), Aladag et al. (2013), Fener and Yesiller (2013), Ozcoban et al. (2018), Tahasildar et al. (2018), Uysal (2020)
(Ashayeri and Yasrebi (2009), Najafzadeh et al. (2013), Abbaszadeh Afshar et al. (2016), Ziaie Moayed et al. (2018), Azan and Haddad (2019), Mohammadi et al. (2019), Zeraatpisheh et al. (2019), Akbarimehr et al. (2020), Naeini et al. (2021)
(Najjar et al. 1996), Dathe and Thullner (2005), Liu et al. (2005), Doris et al. (2008), Kaya (2009), Chittoori and Puppala (2011), Pachepsky and Park (2015), Congress and Puppala (2020)
Bahmed et al. (2019), Dounane and Trouzine (2020), Choudhury and Costa (2019), Costa et al. (2008), Vogel et al. (2005), Kim and Park (2012), Kwak et al. (2020),
A. Abbaszadeh Shahri (2016a, b), Malehmir et al. (2013), Obrzud et al. (2012, 2011), Abdalla et al. (2015), Basma et al. (2003), Aljanabi et al. (2018), Bekhor and Livneh (2014), Botula et al. (2015), Katuwal et al. (2020), Ouf (2012), Ermias and Vishal (2020), Plé et al. (2013), Jalal et al. (2021), Thoang and Giao (2015), Souissi and Trabelsi (2018)
4.3 Stage 2: Topical Analysis
A network of terms, using systematic bibliometric analysis, identifies interests in a specific research domain (Linnemeier et al. 2010; Darko et al. 2020). This scientometric analysis implemented the term co-occurrence, which evaluated the incidence and frequency of two unique terms used in the title, abstract, and keywords sections. This technique identifies the similar terms used in studies utilising AI in reactive soil research and determines how they are organised and interrelated (van Eck and Waltman 2014).
The co-occurrence analysis was performed using VOSviewer® 1.6.17. Full counting was used where all occurrences of terms have been considered. The network of terms or nodes was weighted using a similarity measure known as the association by strength or edge weight (Eck and Waltman 2009). This means the strength of the relation between two nodes is calculated based on the number of papers in which the terms occur together calculated as
$$s_{ij} = \frac{{c_{ij} }}{{w_{i} w_{j} }}$$
(1)
where sij is the similarity between i and j, cij is the number of co-occurrences of nodes i and j, wi and wj are the total number of occurrences/co-occurrences of terms i and j. This calculation of sij influences the node and distance between nodes.
The clusters of research interests using a co-occurrence network of terms are presented in Fig. 5b. The network of terms is categorised into four clusters. These clusters are (1) reactive soil characterisation and swell prediction in red, (2) soil and structure performance in blue, (3) clay cracking and desiccation in yellow, and (4) reactive soil property classification and behaviour in green. There are a total of 41 terms in the network. The most frequent occurrences of terms were “expansive soil” (118), “Artificial Neural Network (ANN)” (72), “parameter” (58), and “prediction” (45). The highest relevance scores are for terms “expansion” (3.5), “grade” (3.2), “uplift capacity” (3.0), “injection depth” (3.0), “information technology” (2.8), and “support vector machine (SVM)” (1.9). Terms with low relevance scores, approximately < 0.10 from van Eck and Waltman (2014) follow a random and less significant pattern, whilst terms with high relevance scores (approximately > 0.10 from van Eck and Waltman 2014) co-occur mainly with limited terms. This means that terms with higher relevance scores represent a significant interest in the research domain.
Fig. 5
Research interest clusters attained using co-occurrence network of terms for a geotechnical engineering keywords from 1981 to 2020, having five clusters, and b reactive soil keywords from 1993 to 2020, having four clusters, created using VOSviewer®
The relative sizes of the nodes and terms in Fig. 5b represent the weights based on occurrences. It is observed that the terms “expansive soil”, “ANN”, “data”, “prediction”, “pressure”, “technique”, “technology”, “crack”, and “strength” are the terms with larger weights. The line connecting two terms reveals that they have appeared together, and the thicker the line the more recorded co-occurrence. In that, nodes are closer to each other if there is a stronger relationship (Liao et al. 2018).
A more detailed clustering considering the temporal aspect of publications was created using Citespace®, as shown in Fig. 6. The clustering has 19 topics that are related to each other. Soft computing techniques (e.g., AI algorithms, ANN, SVM, computational intelligence) and preloading methods have been the largest clusters in the existing literature. These two are the most commonly employed AI techniques for reactive soil investigations. Figure 6 also shows that in 1993, the application of image analysis of porous media to investigate cracking of reactive soil by Hallaire (1993) has been linked and extended to microtomography, propagation algorithm, piezocone test, and geometrical structure studies.
Fig. 6
Detailed clustering structure of studies applying AI to reactive soils with temporal consideration
The clusters shown in Fig. 6 reflect the rapid increase in the publications considering the application of AI in reactive soil. As with Figs. 1 and 6, the increase in term occurrences and linkage of the clusters are again observed between 2001 and 2021. This is in line with the graph in Fig. 3 and the geolocation analysis in Fig. 5b. The most recent activities are from research clusters focusing on landslide, clayey soil, unsaturated compacted soil, microtomography, laboratory study, propagation algorithm, geophysical technique, and piezocone test.
4.4 Stage 3: Network Analysis
The network analysis using bibliometric coupling has been implemented in this scientometric study due to its applicability to short-term analyses and the scientific research domain (Klavans and Boyack 2017). The result of the bibliographic coupling is shown in Fig. 7a. The bibliographic coupling resulted in nine main clusters of bibliographic coupling using authors. The following authors are grouped based on their research application and linked due to possible collaboration or study interests. These nine clusters created through bibliographic coupling is related to the term co-occurrence analysis in Fig. 5b by the term occurrence closely aligning to the authors. The clusters are discussed as follows:
Cluster 1 focused on the implementation of soil treatment and soil stabilisation primarily using fuzzy logic (Ikeagwuani and Nwonu 2020; Darikandeh and Phanikumar 2021). This topic, where AI has been applied, was conducted primarily in Algeria, Canada, China, Egypt, India, Iran, Nigeria, and Turkey (Table 1).
Relatively close to cluster 1 is cluster 2, focusing on soil properties, failure and collapse that also implemented fuzzy logic algorithms (Gong et al. 2004; Ding et al. 2006, 2009; Yin et al. 2018; Yu et al. 2020; Peng et al. 2021; Zheng et al. 2021).
Cluster 3 investigated reactive soil property prediction, geo-structural health monitoring, and design optimisation using ANN, genetic algorithms (GA), and image analysis. Cluster 3 has proximity to clusters 6, 7, 8, and 9, which means these clusters have strong relation due to similar research interests, comparable employed methodologies (e.g., ANN, GA, SVM) and possible study collaboration (Sabat 2015; Mozumder et al. 2017; Pham et al. 2018; Borthakur and Dey 2020; Liu et al. 2020; Congress and Puppala 2020; Naeini et al. 2021; Onyelowe et al. 2021; Jalal et al. 2021).
The remaining clusters 4 and 5 investigated soil cracking and desiccation primarily using image analysis (Costa et al. 2008; Sima et al. 2014; Rasa et al. 2018; Choudhury and Costa 2019; Heikkinen et al. 2019; Ding et al. 2020; Hashim and Sayl 2020).
Fig. 7
Network analysis implementing bibliographic coupling using unit analysis of a authors and b countries
The network analysis by bibliographic coupling using countries and average year of publication is presented in Fig. 7b. In line with the geospatial analysis presented in Fig. 4, the bibliographic coupling in Fig. 7b shows the scientific collaboration network. This helps identify the linkage of countries actively researching the application of AI to reactive soils (van Eck and Waltman 2014; Liao et al. 2018; Darko et al. 2020). VOSviewer® 1.6.17 was used, and a total of 32 countries were identified. The weights used were the total number of research publications that affected the node size and term size for each country in Fig. 7b. Similar to the geospatial analysis in Fig. 4, several studies were from China, India, Turkey, Iran, the US, and Canada. These countries have conducted research investigating damage due to shrink-swell reactive soil movement (Fig. 1) (Zou 2015). Recent research activities and publications in this specific field are observed in China, India, Iran, Iraq, Malaysia, Vietnam, Nigeria, Pakistan, and Uganda (coloured yellow-green to yellow in Fig. 7b). A strong paired relations of countries was from China–US, China–Pakistan, China–Canada, China–India, India–Vietnam, India–Turkey, US–South Korea, Canada–Algeria, Iran–Sweden, and Iran–Italy represented by thicker line connections. Noticeable strong knowledge exchange and collaboration relationships are between Canada, China, India, Turkey, and the US. These countries have also published several articles applying AI to reactive soils.
It is important to note that clusters are generally closer when the relationship is stronger, and thicker lines connecting two terms represent more recorded bibliographic coupling, as proposed by Liao et al. (2018). It can be observed that Clusters 3, 6, 7, 8, and 9 are close to each other, reflecting that their topics or methodologies were comparable or had a connection. Moreover, thicker connecting lines can be seen among authors in Clusters 3, 6, 7, and 9, reflecting that these authors were cited by the same publications. This may indicate that these cited authors may be collaborators or have a comparable research focus.
The bibliographic coupling using authors and countries, in Fig. 7, has shown useful information about the research dynamics of studies applying AI to further understand reactive soils. The research interests, knowledge sharing and collaboration network can be extended to countries having design and construction challenges due to shrink-swell movements. These include countries in North America (the United States of America and Canada), South America (Mexico, Brazil), Europe (the United Kingdom), Africa (Egypt), Asia (Indonesia, Malaysia, Myanmar), and Australia, as shown in Fig. 1.
4.5 Stage 4: Citation Burst Analysis
Citation burst analysis illustrates the most active terms and cited journals in a specific date range (Chen 2006). This technique detects publication events associated with a surge of citations and fast-growing topics (Zhou et al. 2019). This study implemented burst analyses for both keywords (Fig. 8) and cited journals (Fig. 9) using Citespace®. This provides details of topics and publishers that have been considered in the growing interest of applying AI to reactive soils. The cyan-coloured lines in Figs. 8 and 9 denote the scientometric survey date range from 1993 to 2021. The orange-coloured lines represent the length of a citation burst event based on keywords and cited journals. Note that the term co-occurrence analysis in Figs. 5b and 6 is based on degree centrality values, which consider the linkage of nodes in a network. The citation burst analysis, unlike the term co-occurrence analysis, is based on the degree of attention a term, keyword or topic received from the scientific community concerning a specific period. Thus, terms in Fig. 8 may not necessarily be present in Figs. 5b and 6.
Fig. 8
Top 20 keywords or terms with the strongest citation burst in the research domain applying AI to reactive soil investigations from 1993 to 2021. The orange-coloured lines represent the length of a citation burst event
Top 10 cited journals with the strongest citation burst in the research domain applying AI to reactive soil investigations from 1993 to 2021. The orange-coloured lines represent the length of a citation burst event
The top 20 trending topics from 1993 to 2021 are presented in Fig. 8. The first term citation burst represented by orange lines was for “image analysis” in 1993. It can be observed that in the detailed clustering of Fig. 6, the term was “porous media” instead. This difference in terms shown in Figs. 6 and 8 is due to the difference in the analysis as explained in the previous paragraph (degree of centrality values against the degree of attention from the research community). This difference reveals that “image analysis” acquired substantial attention within this specific period but had a lack of application or linkage to other studies related to the application of AI to reactive soil (Darko et al. 2020). For instance, “image analysis” was frequently employed in a few areas (e.g., porous media, desiccation) but was not extended to sufficient applications linking to wider networks (e.g., soil treatment, parameter prediction). Alternatively, the citation burst of “image analysis” can also be directly related or referring to the term cluster “porous media” (Hallaire 1993). The term citation burst also captured the halted publication of AI to reactive soils, where the next term citation burst was experienced in 2002 for “soil pore” and “clay soil”. This paused research activity in this particular field captured in Fig. 8 is in line with the findings in Figs. 3 and 6. From 2002 onwards, citation bursts have been experienced. Relatively strong citation bursts in the study period were “neural network” (2.40 in 2009), “adaptive regression” (1.97 in 2011), “multivariate adaptive regression” (1.97 in 2011), “piezocone test” (1.97 in 2011), “expansive clay” (1.97 in 2013), “California Bearing Ratio (CBR)” (1.97 in 2013), and “two-layered” (2.51 in 2019).
The top 10 most cited journals from 1993 to 2021 are presented in Fig. 9. Citation bursts in journals related to AI and reactive soils, represented by orange lines, were not detected until 2018. The citation burst has coincided since then, between 2018 and 2021. Strong citation bursts were from the Canadian Geotechnical Journal (12.39), Geoderma (10.24), Soil Science Society American Journal (9.66), Soil Science Society of America (9.66), Catena (8.18), and Computer and Geotechnics (7.96). The strong bursts imply that these journals are the trending sources of studies implementing AI in reactive soil investigations.
5 Discussion and Future Direction
The application of AI in the investigations on reactive soils is currently limited, even though different machine learning algorithms have been used in geotechnical engineering. However, there has been a growing interest in this research domain. This study thoroughly surveyed the literature in the past 28 years using a scientometric approach, aiming to evaluate the development of AI methods applied to further understand the characteristics and behaviour of reactive soils.
Current studies employing AI in reactive soils studies are focusing on (1) reactive soil characterisation and strength prediction, (2) soil and structure performance, (3) clay cracking and desiccation, and (4) reactive soil movement and stabilisation (as demonstrated in Fig. 5b). Primary AI techniques used in reactive soil studies were ANN, SVM, fuzzy logic, GA, and image analysis (as presented in Figs. 7, 8, and 9). With the rise of innovative technologies in data availability, storage, and computing, future opportunities in this specialised field of AI utilised to study reactive soils can be extended to (1) microstructure and nanotechnology studies, (2) real-time geo-structural health monitoring, (3) multi-disciplinary forecasting, (4) knowledge sharing and database, and (5) geo-structural design enhancement, presented in Fig. 10.
Fig. 10
Current and future application of AI in studies related to reactive soil
Image analysis has been one of the most applied AI techniques to study reactive soils as observed in Figs. 6, 8, and 9 (Hallaire 1993). This technique is commonly utilised for investigating porous media structure and desiccation. A more specific type of image analysis currently gaining interest is CNN, an edge detection technique commonly implemented in facial recognition, picture classification, document digitisation, and other image-based applications (Shin et al. 2016). CNN has been employed in different applications in geotechnical engineering, mainly focusing on particle microstructure and geo-structural health monitoring (e.g., abutments, tunnels, pavements) (Abdel-Qader et al. 2003; Catbas et al. 2012; Adhikari et al. 2014; Koch et al. 2015; Dorafshan et al. 2019). However, this technique has not been used until now to determine or predict reactive soil properties and behaviour. CNN can be implemented with the incorporation of DL which leads to the increased accuracy of models and analyses.
In the AEC industry, a growing number of studies explore microstructures of cement-based, steel, and timber materials (Kusiak and Kuziak 2002; Jiahe et al. 2003; Cao and Zhang 2004). The increase in the application of different nanomaterials has necessitated the investigation at micro-levels related to physical deterioration, chemical deterioration, reinforcement corrosion, nano-admixtures, and cement hydration. Research studies in the application of AI to microstructure can be classified depending on the materials. The classification, considering the reviewed literature, is divided into the common materials used in structural design and these are cement-based materials, steel, and timber. However, these materials are easier to analyse compared to anisotropic types with a saturation-varying porous material like shrink-swell soils (Tran et al. 2021).
CNN, when paired with tomography, could be a powerful technique to gain insightful knowledge on the behaviour of reactive soils. Few studies have investigated the microstructure using tomography, a radiologic technique for obtaining clear X-ray images (Hyväluoma et al. 2012; Julina and Thyagaraj 2019; Yu et al. 2020). SEM is a tomographic technique that can offer comprehensive topographical and compositional information on reactive soil and its clay content (Liu et al. 2005). Through CNN, DL, and SEM, image recognition of specific patterns in micro to nano-level would be possible that can help determine accurate reactive soil properties (e.g., hydraulic conductivity, shrinking and swelling potential, and detailed characterisation). Using these combined techniques will also be beneficial to the investigations on the strength prediction, desiccation, ground movement, and the stabilisation of reactive soils.
5.2 Real-Time Geo-structural Health Monitoring
AI techniques can be applied to monitor geohazards induced by reactive soils, such as ground movement monitoring and lightweight infrastructure damage (e.g., pipes and pavements) (Rajeev and Kodikara 2011; Teodosio et al. 2020a, b). The current AI implementation in the AEC industry commonly uses a single method to analyse features collected from sensors, for instance, CNN for image analysis, which is termed unimodal analysis. More accurate geo-structural health monitoring could require varying methods or multimodal analysis. For instance, CNN is used for image analysis and recurrent neural network (RNN) is used for sound analysis involving a video dataset. Few studies have integrated video and seismic sensor monitoring to monitor transportation infrastructure, combining computer-vision algorithms (e.g., CNN) and vibration analysis (Gandhi et al. 2007). A similar multimodal application was employed in monitoring bridges and their abutments to prevent failure (Khan et al. 2016). A growing number of multimodal implementations has been observed combining RFID, tag antennas and sensors for multimodal geo-structural health monitoring applications (Zhang et al. 2017).
Another potential application of AI to alleviate reactive soil challenges would be developing and adopting real-time geo-structural monitoring and online learning. With the emergence of IoT, smart cities, and larger datasets, adaptive and time-iterative algorithms can be analysed using a larger volume of data (Theodoridis 2020). Moreover, applying principal component analysis (PCA) to reduce the dimensionality of a dataset whilst preserving statistical information and variability as much as possible (Cleverly et al., 2020) can make real-time geo-structural health monitoring plausible.
To have an automated warning alert, anomaly detection that detects outliers based on historical time-series data can be employed in monitoring systems (Ahmad et al. 2017; Pang et al. 2021). This implies that we can also have safer cities by employing real-time unsupervised anomaly detection for streamed geotechnical, climatological, and structural data giving hazard or risk warnings when an inconsistency is sensed in a dataset (Ahmad et al. 2017). These combined techniques will be most beneficial to shrink-swell movement monitoring to minimise lightweight structure damage and disaster alert (e.g., mudslide or long-term drought).
5.3 Multi-disciplinary Forecasting
Remote sensing geohazards due to changing climate is necessary to reduce the potential damage and disaster (Pritchard et al. 2015). Issues related to reactive soils require a multi-disciplinary approach incorporating geotechnical, environmental, climatological, and structural knowledge domains (Teodosio et al. 2021). Thus, the complex challenges brought by reactive soils require multi-physical non-linear analysis for heterogeneous and anisotropic material (Shao 2017). Furthermore, it is important to consider where issues related to reactive soil are occurring, as shown in Figs. 5 and 8b.
Geospatial analysis and remote sensing can be used to determine the short-term and long-term impacts of climate change on the seasonal volume variation of reactive soils. By combining AI and remote sensing techniques, the insights will be placed into the context and can consider spatiotemporal changes of essential variables. For instance, more extreme rainfall events and prolonged droughts can cause severe geohazard and infrastructure damage due to a wider range of soil suction changes in reactive soils (Leao 2014; Trenberth et al. 2014; Elguindi et al. 2014; Teodosio 2017; Teodosio et al. 2017; Rhee and Im 2017; Varquez and Kanda 2018).
The geohazard mapping due to reactive soils can be worthwhile to improve the asset management of infrastructure and dwellings, considering the effect of climate change (Pritchard et al. 2015). Employing AI and remote sensing together can also reduce disaster occurrence and give context and more insights to obtain results.
5.4 Knowledge Sharing and Global Database
The lack of knowledge-sharing and research collaboration have been identified using the bibliographic coupling network analysis in Fig. 7b. Developing a knowledge-sharing database can provide standardised pre-processing and post-processing methods (i.e., quality control and gap-filling using machine learning) across the network for a better spatiotemporal comparison of sites monitored or analysed (Isaac et al. 2017). This will also help scientists or researchers focus on the analysis and interpretation of data. Moreover, data and site information (i.e., metadata) can have a standard file format available from online portals (Nelson et al. 2020). This will encourage knowledge sharing and expedite innovations through the standardised application of AI in reactive soil studies.
5.5 Geo-structural Design Enhancement
The application of AI could be beneficial in creating and understanding the novel and sustainable materials and geo-structural design. Enhancements can be approached in two different ways, either through geotechnical engineering or structural engineering.
An example of implementing AI to enhance design through geotechnical engineering is soil stabilisation using recycled materials employing circular economy concepts (e.g., glass, tyres, natural fibres, crushed ceramics and concrete) (Al-Bared et al. 2018; Arrieta Baldovino et al. 2020; Al-Baidhani and Al-Taie 2020; Yaghoubi et al. 2021). A circular economy is defined as a renewing scheme where input and waste materials, emissions, and energy outflow are minimised by decelerating, terminating, and lessening material and energy loops (Geissdoerfer et al. 2017). Employing AI to determine the optimum mix proportions and stabilisation depth through parametric experimental data can lead to more sustainable solutions.
Considering the structural engineering aspect, incorporating AI into designing lightweight structures such as pavements and residential foundations based on empirical data or parametric simulations can lead to more accurate calculations by capturing non-linear behaviour relationships between features (Solanki et al. 2009). Another innovative solution is the use of biomimetic designs, a concept that incorporates solutions based on nature to structural challenges (Jamei and Vrcelj, 2021). The generative design and AI have been used in biomimetic designs to produce optimised materials and structures (Challapalli and Li 2025). This can be utilised in lightweight infrastructures and low-rise residentials prone to damage due to shrink-swell ground movements.
Based on the literature analysis, the probable research avenues implementing various AI techniques have been shown. With the inception of interconnected wireless technologies, environmental-conscious design, and innovative solutions, the emerging fields of AI application to reactive soil studies can lead to microstructure and nanotechnology investigations, real-time geo-structural health monitoring, multi-disciplinary forecasting, knowledge sharing and databases, and geo-structural design enhancements.
AI algorithms that were used in different civil engineering applications are listed in Table 3. These can have emerging applications that can be useful in investigating and resolving challenges in the reactive soil research (Table 3).
Table 3
AI algorithms that can have emerging applications to reactive soil research
Algorithm
Advantages
Disadvantages
Possible applications
Reference/s
Convolutional Neural Network (CNN)
Learns relevant spatial features through image analysis
Valuable in landslide detection, land-use classification, and terrain analysis
Large datasets are often required to avoid overfitting
Computationally demanding,
Interpretability can be challenging
Landslide detection
Identifying subsurface features from GPR data
Automated crack detection
Predict the extent of soil expansion and shrinkage
Based on the literature analysis, the probable research avenues implementing various AI techniques have been shown. With the inception of interconnected wireless technologies, environmental-conscious design, and innovative solutions, the emerging fields of AI application to reactive soil studies can lead to microstructure and nanotechnology investigations, real-time geo-structural health monitoring, multi-disciplinary forecasting, knowledge sharing and databases, and geo-structural design enhancements.
6 Conclusion
This manuscript explores the application of Artificial Intelligence (AI) in reactive soil research, particularly in the context of understanding the impact of reactive soils on physical infrastructure such as roads and residential structures. Over the last 35 years, AI has been successfully utilised in geotechnical engineering to process large volumes of data, create adaptive non-linear models, and predict valuable insights. This study conducted a scientometric analysis of research literature from the past 28 years to identify key trends, AI techniques, and gaps in current applications to reactive soils.
The leading AI techniques identified through co-occurrence network analysis include artificial neural networks (ANN), support vector machines (SVM), fuzzy logic, genetic algorithms (GA), and image analysis. These methods have been primarily applied to soil characterisation, strength prediction, soil-structure interaction, clay cracking and desiccation, and soil movement and stabilisation. However, gaps in the research were identified, particularly the lack of application of advanced AI techniques that are emerging in industries like finance, automotive, telecommunications, and manufacturing to shrink-swell ground movement studies.
There is a need to integrate more advanced AI methods, such as deep learning and reinforcement learning, which have proven successful in other sectors but remain underexplored in reactive soil studies. These techniques could improve predictive modelling accuracy and uncover hidden patterns in complex soil behaviours.
Future studies should explore the integration of AI with emerging technologies in fields like microstructure analysis, nanotechnology, and real-time geo-structural health monitoring. Combining these interdisciplinary approaches could lead to more accurate models for soil behaviour prediction and infrastructure performance.
A major challenge in applying AI to reactive soil research is the "black box" effect, where complex AI models are not easily interpretable. To make AI models more practical and reliable for geotechnical engineers, there should be a focus on improving algorithm transparency and interpretability, ensuring that decision-makers can confidently apply AI-driven solutions in real-world scenarios.
As AI technologies evolve, so too will their applications in reactive soil research. The emergence of interconnected wireless technologies, sustainable design principles, and innovative solutions could open new research opportunities. Future work could include investigating the potential of AI for real-time monitoring of soil conditions, developing multi-disciplinary forecasting models that incorporate both soil and structural data, and creating knowledge-sharing databases for global collaboration. Additionally, research could focus on the use of AI for geo-structural design enhancements and optimisation of soil stabilisation techniques.
Overall, while AI has already made significant strides in reactive soil research, there is immense potential for further development. Addressing current gaps and focusing on the application of advanced AI techniques, interdisciplinary approaches, and algorithm transparency will help drive the field forward, making AI a more valuable tool for predicting and mitigating the impacts of reactive soils on infrastructure.
Acknowledgements
This work was funded in partnership with the Victorian State Government that the authors would like to acknowledge and thank.
Declaration
Competing interests
No potential conflict of interest was reported by the authors.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
