State-of-the-Art Review of Artificial Intelligence in Environmental Geophysics and Geotechnical Engineering

1. Introduction

Artificial intelligence (AI) is reshaping environmental geophysics and geotechnical engineering (EGGE) through the AI–EGGE paradigm, redefining how near-surface systems are characterized, modeled, and managed under increasing environmental and infrastructure pressures [1,2,3,4]. While conventional and physics-based approaches remain foundational, they often struggle to represent the nonlinear behavior, uncertainty, and multiscale heterogeneity that govern surface–subsurface processes [5,6,7]. The rapid growth of geophysical, geotechnical, and environmental datasets has positioned AI as a powerful complement, enabling the extraction of complex patterns from high-dimensional data and improving inference where traditional inversion and regression are poorly constrained [8,9]. By learning relationships among geophysical observables, mechanical properties, and environmental indicators, AI enhances prediction, interpretability, and decision support in subsurface characterization [10,11,12].

The adoption of AI in Earth sciences has progressed from early expert systems and regression-based models to modern machine learning (ML), including unsupervised clustering algorithms such as k-means, fuzzy c-means, and self-organizing maps, capable of extracting spatial, temporal, and structural features from complex datasets [13,14,15,16]. These foundations have further advanced into deep learning (DL) architectures, e.g., feedforward, convolutional, recurrent, autoencoder, generative, and graph-based networks, that autonomously learn hierarchical representations from high-dimensional data [17,18,19,20]. More recently, hybrid and physics-informed approaches have emerged, explicitly embedding governing physical constraints within data-driven learning to support multi-physics integration across geophysical and geotechnical domains [21,22,23]. These developments now enable coupled modeling of electrical (resistivity, induced polarization [IP], self-potential [SP]), electromagnetic (EM), seismic (refraction tomography and multichannel analysis of surface waves [SRT, MASW], reflection), ground-penetrating radar (GPR), magnetic, gravity, radiometric, and borehole datasets within unified AI-driven predictive frameworks.

EGGE are inherently complementary disciplines focused on imaging, characterizing, and engineering the near-surface critical zone, where hydrological, geological, geochemical, and geomechanical processes interact [24,25]. Environmental geophysics provides noninvasive, continuous subsurface imaging through the aforementioned geophysical datasets, enabling delineation of subsurface structure, fractures, contaminant pathways, voids, cavities, hydrogeological boundaries, etc [26,27,28]. Geotechnical engineering complements these observations through direct characterization of soil and rock behavior using the standard penetration test (SPT/SPT-N), cone penetration test (CPT/CPT-qc), rock quality designation (RQD), rock mass quality (RMQ), and related indices [29,30]. Integrating these heterogeneous datasets within AI-enabled workflows, hereafter termed AI–EGGE, unites continuous subsurface imaging with discrete mechanical and material properties, enabling physics-consistent feature extraction, improved prediction of geotechnical indices from geophysical attributes, and enhanced spatial resolution of geomechanical parameters.

Applications of AI–EGGE are expanding rapidly across surface–subsurface and infrastructure systems. Supervised and ensemble models have been used to predict shear-wave velocity (Vs) and stiffness from ERT, MASW, or SRT [6,21,31], and infer P-wave velocity (Vp), dynamic moduli, and soil or rock integrity using ERT–SRT or ERT–IP correlations [13,27,32,33,34]. Unsupervised ML and DL further improve mapping of subsurface heterogeneity, facies, transition zones, and landslide precursors from integrated geophysical–borehole data [35,36,37]. Hybrid and physics-informed models, including Bayesian and metaheuristic–DL couplings, enhance estimation of integrity, deformation moduli, and stress–strain behavior for slope, tunnel, and foundation design [38,39]. Cross-disciplinary DL architectures, such as encoder–decoders, convolutional neural networks (CNNs), U-Nets, vision transformers (ViTFs), and graph neural networks (GNNs), extend AI–EGGE to multi-source spatiotemporal modeling [18,40,41]. By integrating gravity, thermal, and hydrogeological indicators, these models support improved contaminant tracking, groundwater dynamics, deformation monitoring, and resilience planning [42,43]. Collectively, they are advancing intelligent digital-twin–like systems and scenario-based simulation for infrastructure resilience and environmental risk management [3,44].

Given this background, this review consolidates recent theoretical and practical advances in AI–EGGE within a unified workflow framework (Figure 1). Specifically, the review: (i) synthesizes the historical and conceptual development of AI methodologies relevant to EGGE; (ii) examines strategies for data integration, feature engineering, and cross-domain coupling between geophysical and geotechnical datasets; (iii) reviews representative applications and case studies linking geophysical attributes with geotechnical indices; (iv) critically assesses persistent challenges related to data scarcity, generalization, interpretability, and validation; and (v) outlines pathways toward physically grounded, uncertainty-aware, and explainable AI (XAI) systems. Guided by this structure, the paper progresses from methodological foundations to integrated applications and forward-looking perspectives on environmental-engineering-grade AI for near-surface characterization.

2. Environmental Geophysics and Geotechnical Engineering: Foundations of AI–EGGE

The evolution of AI, from early neuron models in the 1940s through the first “golden age” of the 1950s–1960s, subsequent stagnation cycles, and the modern DL era, has progressively driven the convergence of geophysical and geotechnical domains by enabling multimodal data fusion, hierarchical feature learning, and predictive modeling across heterogeneous surface–subsurface datasets [45]. Conventional EGGE workflows, however, struggle to resolve nonlinear interactions and spatiotemporal coupling governing near-surface processes. These limitations motivate AI–EGGE as an integrative paradigm augmenting physics-based methods with data-driven learning for holistic subsurface characterization and resilience-oriented engineering. This Section outlines the core principles, data structures, and methodological foundations underpinning AI–EGGE.

2.1. Environmental Geophysics: Methods, Capabilities, and Integrated Applications

Environmental geophysics employs noninvasive near-surface techniques to characterize subsurface conditions controlling groundwater flow, contamination, land-use planning, and ecosystem resilience [12,46]. Core methods include ERT/IP, GPR, time- and frequency-domain EM, seismic approaches (SRT, MASW), and remote sensing. ERT delineates conductivity contrasts linked to aquifers, saline intrusion, and contaminant plumes but remains sensitive to saturation, temperature, and pore-fluid chemistry [28,36,47]. IP complements ERT by capturing capacitive responses associated with clay content and mineral surface properties, enhancing lithological and contaminant discrimination [27,48]. GPR provides high-resolution imaging in resistive environments yet attenuates in conductive media [25,49], while EM offers rapid coverage at lower vertical resolution [47,50]. Seismic methods (Vs, Vp) yield elastic, stiffness, and rippability, strengthening interpretation when integrated with electrical properties [13,33]. Remote sensing supplies regional constraints on land use, vegetation stress, and recharge variability [51,52]. Multi-method integration, joint inversion, and AI-driven fusion reduce non-uniqueness, improve uncertainty quantification (UQ), and advance physics-informed multi-physics and ML/DL frameworks [4,9].

Computational advances and autonomous sensing technologies are accelerating operational deployment [53,54]. Two dominant trends emerge: (1) fusion of geophysical and Earth observation (EO) datasets using statistical, geostatistical, and ML/DL frameworks for time-lapse characterization and forecasting, and (2) adoption of UAVs, robotic platforms, and dense sensor networks delivering high-spatiotemporal data streams. Multispectral and hyperspectral imagery detect land-surface dynamics, while field geophysics constrains subsurface structure [55,56]. Multi-sensor fusion is increasingly demonstrated in practice: Gholizadeh et al. [57] merged satellite data with ERT for regional soil-contamination mapping; Omolaiye et al. [58] integrated geophysical and remote-sensing datasets with hydrogeological models to predict groundwater availability in a semi-arid Nigerian basin; and Davis et al. [51] reviewed the convergence of sensor-based monitoring, computational techniques, and Internet of Things (IoT) for soil and groundwater contamination assessment. These hybrid approaches support integrated hydro-environmental modeling and watershed-scale decision-making. Physics-informed ML, joint inversion, and cloud-based processing further reduce interpretive ambiguity, enable transfer learning under data scarcity, and support adaptive survey design. The convergence of autonomous sensing, AI-driven processing, and cloud orchestration now underpins AI–EGGE, shifting practice from method-centric surveys toward integrated subsurface intelligence [39,59].

2.2. Geotechnical Engineering: Testing, Instabilities, Monitoring, and Computational Integration

Geotechnical engineering focuses on the mechanical and hydraulic behavior of soils and rocks under natural and engineered loading. Conventional in-situ and laboratory tests—SPT, CPT, RQD/RMQ, consolidation and shear testing, seismic profiling, and electrical methods—remain central to evaluating bearing capacity, settlement, and slope stability [14,30]. While deterministic datasets historically supported empirical correlations and factor-of-safety approaches, sensitivity to heterogeneity and uncertainty has driven the adoption of probabilistic and performance-based frameworks [60,61]. Advanced seismic testing constrains dynamic properties essential for performance-based design, while reliability methods propagate parameter uncertainty through safety and serviceability assessments [62,63]. These developments reinforce the need for integrated multi-test characterization.

Advances in non-destructive testing (NDT) have expanded the geotechnical toolkit [64]. GPR, ERT, and IP enable high-resolution subsurface imaging without disturbance, enhancing data triangulation when combined with conventional tests. Probabilistic slope-stability modeling, spatial random fields, and Bayesian updating now support quantified, scenario-based risk assessment for design and mitigation [65], moving practice beyond deterministic parameters. NDT and time-lapse geophysical monitoring also capture construction effects and seasonal hydromechanical variability [26,66]. These capabilities support improved foundation design, slope-stability evaluation, and excavation planning [60,67]. Distributed fiber-optic sensing (DFOS), including Brillouin-based methods, optical time-domain reflectometry, distributed temperature sensing, and distributed acoustic sensing (DAS), provides dense, continuous measurements of strain, temperature, and acoustic response along structural elements and slopes [68,69]. These streams enable early detection of deformation and instability precursors, although installation, calibration, and data-management challenges persist [70]. Coupled with AI-driven anomaly detection and explainable ML, DFOS supports adaptive early-warning systems.

Climate-driven variability within the near-surface critical zone is redefining boundary conditions and increasing service-life uncertainty, necessitating coupled hydrometeorological–geotechnical models [51,71]. AI and optimization approaches increasingly support scenario-based slope-failure assessment under future climatic pathways by integrating testing, monitoring, and physical modeling, which are central to computational geotechnics [8,72]. These enable the prediction of soil classification, shear strength, liquefaction potential, bearing capacity, and rock-mass indices [6,73,74]. Ensemble learners, e.g., random forest (RF), perform robustly under noisy, heterogeneous datasets [75], while DL captures nonlinear and spatial dependencies when adequately trained [76,77]. Coupling probabilistic modeling with optimization techniques such as particle swarm optimization (PSO) and genetic algorithms (GAs) improves slope-stability analysis through enhanced parameter calibration and sensitivity assessment [78]. Statistical and data-driven analytics further improve soil behavior interpretation under varying loading and environmental conditions [6,42]. Concerns regarding generalization and interpretability are accelerating hybrid and physics-informed ML frameworks embedding constitutive constraints [11,41].

2.3. Cross-Domain Synthesis in AI–EGGE

AI–EGGE paradigms integrate environmental geophysics and geotechnical engineering within a unified intelligence-driven decision-support framework anchored on three pillars: (i) observational layers (multi-method geophysics, geotechnical testing, DFOS), (ii) computational cores (joint inversion, probabilistic inference, physics-informed/hybrid ML/DL, metaheuristic optimization), and (iii) operational outputs (forecasting, anomaly detection, early warning, decision support). Convergence emerges when geophysical, EO, and monitoring datasets are fused with well logs, CPT/SPT, and DFOS into harmonized feature spaces powering physics–ML hybrid models. These systems enhance subsurface interpolation, contaminant-pathway mapping, slope-instability forecasting, and foundation assessment under transient hydro-mechanical and climatic forcing. Three frontiers accelerate this transition: (a) multi-physics joint inversion reducing ambiguity and improving UQ relative to sequential workflows; (b) physics-informed ML embedding conservation and constitutive laws to preserve realism under sparse data; and (c) cloud–edge infrastructures enabling real-time ingestion, adaptive updating, and automated alerts. Collectively, these advances shift practice toward continuous, intelligence-driven subsurface monitoring, the hallmark of AI–EGGE.

Operationalization depends on interpretability, data integrity, and stakeholder trust. XAI and UQ remain essential where outputs inform safety-critical or regulatory decisions, such as contaminant-transport risk to water sources, foundation reliability assessments, or slope failure warnings in inhabited areas [79,80,81]. Current research highlights the need for transparent hybrid models (e.g., surrogate modeling with uncertainty bounds), standardized data-provenance protocols, and harmonized sensor-calibration and inversion workflows. Socio-technical factors, such as capacity building, data governance, interoperability, and regulatory alignment [82,83], will ultimately govern AI–EGGE’s real-world impact. Establishing robust, traceable pipelines from sensing to decision-making is, therefore, as critical as improving predictive accuracy.

3. AI Techniques in EGGE

AI enables capture of nonlinear, heterogeneous, and spatiotemporally variable subsurface relationships that physics-only methods struggle to resolve, especially under data scarcity [10,84]. Multimodal data fusion strengthens AI–EGGE by harmonizing information from diverse sensing and testing modalities into a unified analytical framework for subsurface characterization. Figure 2 summarizes the three main fusion strategies—early, middle, and late—and their roles in enhancing robustness, generalization, and interpretability across supervised, unsupervised, hybrid, and physics-informed ML/DL workflows. Early fusion concatenates calibrated inputs (e.g., resistivity, Vp, Vs, SPT-N, RQD) into a single feature space; middle fusion integrates learned representations; and late fusion combines modality-specific predictions via averaging, stacking, or uncertainty-weighted ensembling. Early fusion suits co-registered data, middle fusion handles heterogeneous inputs, and late fusion is preferred when modalities differ and uncertainty must be addressed. AI-driven fusion improves efficiency and sustainability by reducing investigation time and cost while enhancing spatial coverage and interpretational consistency.

Supervised, clustering, and ensemble ML methods now link geophysical attributes with mechanical, hydrogeological, and geochemical soil–rock properties, improving spatial coverage, consistency, and investigation efficiency. DL architectures, including deep neural networks (DNNs), CNNs, recurrent neural networks (RNNs), long short-term memory (LSTMs), autoencoders (AEs), transformers (TFs), generative algorithms (GAs), generative adversarial networks (GANs), and physics-informed neural networks, further extract latent features from multimodal datasets for high-resolution 2D/3D subsurface modeling, temporal analysis, anomaly detection, contaminant mapping, and deformation forecasting [10]. AI-driven analytics integrate multi-temporal satellite, geophysical, and sensor data to assess contamination, groundwater vulnerability, and soil degradation, enabling early warning and decision support [72]. Increasing engineering complexity has accelerated hybrid models that fuse geophysical insight with geotechnical parameterization. Feedforward neural networks, especially multilayer perceptrons (MLPs), remain effective for regression and classification of subsurface properties [85]. Recent advances include GAN-based synthetic data generation for inversion and physics-informed models that embed governing constraints to improve generalization and physical consistency [30].

3.1. Supervised and Unsupervised Learning in AI–EGGE

Supervised and unsupervised learning underpin AI–EGGE by enabling automated prediction, pattern discovery, subsurface characterization, and decision support across integrated geophysical–geotechnical investigations. They are widely applied to model and predict Vs, Vp, elastic moduli, SPT-N and CPT-qc values, compressive strength, bedrock depth, fracture density, porosity, weathering grade, and geohazard susceptibility using geophysical, remote sensing, and geospatial methods.

3.1.1. Supervised Learning

Supervised learning establishes quantitative relationships between input variables and labeled outputs, enabling the estimation of key geotechnical and subsurface integrity parameters from surface and borehole datasets [78,86,87]. Common approaches include linear, multiple, polynomial, and regularized regression methods—simple linear regression (SLR), multiple linear regression (MLR), least absolute shrinkage and selection operator (LASSO), and Ridge regression—alongside support vector machines (SVMs), k-nearest neighbors (KNN), decision trees (DT), RF, M5 model trees, extremely randomized trees (XRT), gradient boosting machines (GBMs) and variants (GBoost, XGBoost, sGBM, LightGBM), categorical boosting (CatBoost), and artificial neural networks (ANNs), including shallow and deep architectures (SNN, DNN). These models capture linear to highly nonlinear relationships and complex feature interactions, while genetic expression programming (GEP) supports interpretable symbolic regression and equation-based prediction. In general, regression-based methods model linear–polynomial trends; SVMs optimize class separation using kernel-based margins; KNN applies instance-based learning; tree-based and ensemble models enable rule-based decision modeling; boosting frameworks improve sequential ensemble performance; and SNN/DNN learn hierarchical nonlinear mappings through multi-layer feature extraction [13,31,32,42,88]. GEP is particularly beneficial in EGGE as it yields explicit analytical equations that support interpretability, parameter sensitivity assessment, and engineering adoption [24,85].

The models illustrated in Figure 3 summarize key supervised ML approaches widely used in AI–EGGE. The ANN (a) comprises an input layer, one or more hidden layers, and an output layer, where weighted connections and nonlinear activation functions learn complex input–output relationships for regression and classification in geophysical–geotechnical datasets. SVM (b) constructs an optimal separating hyperplane for classification or regression by maximizing the margin between classes, where training samples are represented as vectors $x_i$with corresponding labels $y_i$, and the decision function is defined as $f\left(x\right)=w^{T}x+b$. Kernel functions, $K(x_i,x_j)$ (e.g., linear, polynomial, RBF), map nonlinearly separable data into a higher-dimensional feature space, enabling SVM to handle complex geophysical–geotechnical patterns; support vectors—the samples closest to the margin—control the model’s decision boundary. DT (c) models split data into hierarchical decision rules at internal nodes to minimize impurity or prediction error, with leaves producing final class labels or numeric outputs; DTs are intuitive, transparent, and effective for baseline geotechnical assessments. RF (d) constructs an ensemble of $k$ decision trees using bootstrap sampling and random feature selection, with individual predictions aggregated through majority voting (classification) or averaging (regression) to reduce overfitting and improve generalization. GBM (e) builds trees sequentially, with each new tree fitting the residuals of the previous learner to minimize a chosen loss function; the final prediction is the additive combination of all weak learners $W_1+W_2+\dots +W_n$, enabling strong nonlinear predictive capability. CatBoost (f) is an advanced gradient-boosting algorithm optimized for heterogeneous datasets with categorical variables, transforming categorical features using ordered target statistics to prevent leakage and overfitting, and building symmetric (oblivious) trees at each iteration to enhance training stability, speed, and generalization for diverse environmental/geotechnical predictors.

Model performance is evaluated using R² score, mean absolute error (MAE), mean squared error (MSE), root mean squared error (RMSE), root mean squared logarithmic error (RMSLE), mean absolute percentage error (MAPE), median absolute percentage error (MdAPE), Nash–Sutcliffe Efficiency (NSE), and Willmott’s Index of Agreement, with robustness assessed through repeated k-fold cross-validation, hold-out testing, residual analysis, and bias–variance diagnostics [32,75,89]. XAI techniques such as SHapley Additive exPlanations (SHAP), Local Interpretable Model-Agnostic Explanations (LIME), and Accumulated Local Effects (ALE) are increasingly used to interpret model behavior, quantify feature influence, and enhance transparency and trust in geoscience-focused machine learning applications [9].

3.1.2. Unsupervised Learning

Unsupervised learning reveals inherent structure in unlabeled datasets, enabling lithological zoning, subsurface pattern recognition, facies classification, anomaly detection, hydrogeophysical domain mapping, and structural boundary delineation without ground truth [37,90]. Unlike supervised learning, which relies on error-based metrics, unsupervised model quality is assessed using cluster validity and topology-preservation indices. Clustering is the most widely applied unsupervised approach in AI–EGGE: k-means partitions data into $k$ clusters by minimizing within-cluster variance, while fuzzy c-means assigns membership degrees to better capture transitional boundaries in weathered, fractured, or hydrogeologically gradational zones [91]. As shown in Figure 4, k-means clustering (a) partitions a dataset into $k$ clusters by assigning each input vector $x_j$ to the cluster with the nearest centroid $c_i$ based on a distance metric (commonly Euclidean), then updates centroids iteratively until convergence to reveal latent subsurface patterns [86].

On the other hand, self-organizing maps (SOMs) provide a topology-preserving projection of high-dimensional geophysical attributes onto a 2D neuron lattice, with 3D extensions for enhanced topology retention, supporting visualization of structural trends, material contrasts, and alteration zones [90]. In SOM (b), each neuron ($i$) has an associated weight vector ${\epsilon }_i$. For each input vector $x_j$, neurons compete based on a similarity metric (typically Euclidean distance), and the closest neuron is identified as the Best Matching Unit (BMU) [92]. The BMU and its neighboring neurons are iteratively updated using a neighborhood function and learning rate to preserve topological structure. A Unified Distance Matrix (U-matrix, U) is commonly used to visualize the trained SOM, where low values indicate cluster interiors and high values highlight cluster boundaries. This enables SOM to effectively cluster geophysical–geotechnical signatures for lithological classification, anomaly detection, and environmental pattern recognition.

Advanced variants such as growing self-organizing networks (GSONs) and adaptive resonance theory networks refine cluster boundaries to represent heterogeneous and evolving subsurface terrains [93]. Common clustering and dimensionality-reduction tools include hierarchical clustering for multiscale geological grouping; density-based spatial clustering of applications with noise (DBSCAN) and ordering points to identify the clustering structure (OPTICS) for density-based anomaly detection in irregular data; Gaussian mixture models for probabilistic litho-facies characterization, k-medoids for noise-resilient clustering; principal component analysis (PCA) for feature decorrelation; independent component analysis (ICA) for source isolation; factor analysis for latent structure extraction; and nonlinear manifold learning tools such as t-Distributed Stochastic Neighbor Embedding (t-SNE) and uniform manifold approximation and projection (UMAP) for high-dimensional pattern visualization and structural trend discovery [94,95]. Cluster quality is typically evaluated using internal indices such as Silhouette, Calinski–Harabasz, Davies–Bouldin, Dunn, and S_Dbw, while fuzzy/probabilistic methods employ the Partition Coefficient, Partition Entropy, Xie–Beni, Fuzzy Silhouette, and Kwon indices [96,97]. When partial labels exist, external indices, including Adjusted Rand Index, Normalized Mutual Information, Purity, and Jaccard Similarity, enable benchmarking against known geological or geotechnical classes [98].

AI–EGGE now integrates semi-supervised, self-supervised, hybrid, and physics-informed learning to overcome limited labeled data and improve model generalization. Semi-supervised learning leverages small labeled and large unlabeled datasets through self-training, co-training, and graph-based approaches [87,99]. Self-supervised learning extracts feature representations from unlabeled geophysical data via masked reconstruction and contrastive or patch-prediction tasks [100]. Hybrid workflows combine unsupervised clustering or dimensionality reduction with supervised training to enhance accuracy while preserving geological heterogeneity [25]. Physics-informed learning, including physics-informed machine learning (PIML), physics-informed simple-to-ensemble regressors (PISERs), and physics-informed neural networks (PINNs), embeds governing laws, constitutive relationships, and boundary constraints into model architectures or loss functions to suppress non-physical solutions [100,101], and can be coupled with geostatistics, inversion, and simulation for physical consistency. Model optimization further strengthens AI–EGGE through hyperparameter tuning using Bayesian optimization, evolutionary, and swarm algorithms such as GAs, PSO, differential evolution, Gray Wolf Optimizer (GWO), and hybrid Bayesian optimization–deep learning (BO-DL) strategies for architecture search and loss balancing in physics-informed models [31,100,102]. Together, these approaches deliver more data-efficient, physically consistent, and scalable AI solutions for subsurface characterization, hazard assessment, and infrastructure resilience.

3.2. Deep Learning Approach

Deep learning (DL), a branch of ML built on multi-layer ANNs, has become a dominant framework for high-dimensional analytics by autonomously learning hierarchical, task-specific features from raw or minimally processed data [103,104]. Its adoption within EGGE continues to expand, improving high-fidelity analysis and prediction across geological and structural feature detection [105], landslide forecasting [106], seismic interpretation [107], groundwater modeling [108], infrastructure monitoring [109], soil and rock classification [13,72], geospatial analytics [110], mineral resource evaluation [111], and environmental impact assessment [112].

Building on the ANN backbone, modern DL increases network depth and architectural complexity to enhance representational capacity, abstraction, and generalization. ANNs provide the foundational neural-network paradigm from which contemporary DL architectures are derived, with advanced models, such as DNNs, CNNs, RNNs, LSTMs, AEs, and deep belief networks (DBNs), representing specialized extensions tailored to spatial, temporal, dimensionality-reduction, and hierarchical feature-learning tasks [105,113,114]. Within EGGE, CNNs and AEs support subsurface imaging, signal refinement, inversion enhancement, and lithological/facies classification [115,116]; RNN/LSTMs model time-dependent deformation, settlement, and slope-movement behavior [117,118]; and DNNs model constitutive behavior, stiffness, and strength for engineering design and risk assessment [119]. Generative and uncertainty-aware architectures, including variational autoencoders (VAEs) and GANs for synthetic data augmentation, scenario simulation, and uncertainty representation, and Bayesian neural networks (BNNs) for probabilistic prediction with quantified epistemic and aleatoric uncertainty, are increasingly enabling more robust and risk-informed decision-making [41,113,117]. DL performance is evaluated using standard ML metrics, alongside training–validation loss dynamics, convergence stability, efficiency, and uncertainty diagnostics for probabilistic and generative models.

3.2.1. ANN Architecture

ANNs are widely used in EGGE for classification and prediction, offering a robust alternative to traditional mathematical formulations for modeling nonlinear soil/rock structure interactions [120,121]. The ANN architecture—an input layer, one or more hidden layers, and an output layer linked by weighted neurons—forms the foundation of modern DL architectures [78]. As shown in Figure 5, ANNs learn nonlinear input–output relationships by adjusting synaptic weights during training, enabling the mapping of geoenvironmental and geotechnical variables (e.g., loading, soil parameters, environmental influences) to mechanical responses. Model complexity, driven by the number of hidden layers and neurons, strongly affects accuracy [122]. A key extension is the Fuzzy Inference System (FIS), which uses fuzzy logic and rule-based reasoning to handle uncertainty in EGGE data [42,123]. Its hybrid evolution, the Adaptive Neuro-Fuzzy Inference System (ANFIS), integrates ANN learning with fuzzy logic for nonlinear modeling with improved interpretability [124]. Another classical variant, the Probabilistic Neural Network (PNN), applies Bayesian decision theory with radial basis activation functions for reliable pattern recognition under uncertainty [125]. ANN training commonly uses Backpropagation Neural Network (BPNN), which updates weights via gradient-based optimization to minimize loss, with optimizers such as Adam and RMSProp improving convergence and generalization [123]. Metaheuristic-based training further reduces error in complex applications [56].

Recent ANN-derived architectures: GNNs for spatial-relational learning, TFs for long-range dependency modeling, and PINNs that embed physical laws, are accelerating a shift toward physics-aware DL [18,39]. This promotes domain-informed, interpretable, and more generalizable models that integrate geoscientific and geotechnical priors. Project DeepGeo [62] exemplifies this by embedding geological knowledge into ANN/DNN workflows through structured training-image databases for subsurface characterization in data-scarce settings; 54 expert-interpreted cross-sections enabled Bayesian ensemble learning and stratigraphic UQ, demonstrating the value of knowledge-informed DL. Karpatne et al. [126] likewise showed that ANN-based models outperform empirical and statistical approaches by learning latent features governing stress–strain behavior, stiffness, and strength under variable loading and environmental conditions. Despite progress, limitations—data scarcity, spatial bias, overfitting, weak cross-site generalization, and black-box opacity—still hinder large-scale EGGE deployment. Emerging solutions integrate XAI (e.g., SHAP, LIME, Gradient-weighted Class Activation Mapping [Grad-CAM], uncertainty-aware learning, hybrid physics-informed models, and attention-based architectures to improve reliability, transparency, and engineering adoption. Table 1 further summarizes selected reviews and research on ANN applications.

Figure 5. Typical ANN framework for DL–based prediction of geomechanical properties (after [127]).

Figure 5. Typical ANN framework for DL–based prediction of geomechanical properties (after [127]).

Table 1. Summary of selected reviews and research on ANN with its variants in EGGE.

Authors	Methods used	Results	Relevance	Limitations
[128]	Time Delay Neural Networks (TDNNs) applied to cyclic swelling data from the powerhouse cavern (Iran)	TDNN successfully modeled cyclic swelling pressure with good accuracy, capturing time-dependent behavior	Demonstrates ANN capability in modeling cyclic swelling/shrinkage of weak rocks (mudrock), critical for underground structures	Dependent on site-specific data, generalizability to other rock formations is uncertain
[129]	ANN (MLP, RBF) and multivariate regression with multivariate non-linear regression (MLR, MNR) to predict shear strength parameters in part of Iran	MLP-ANN outperformed RBF-ANN; MLR outperformed nonlinear regression; ANN captured complex nonlinear soil behavior	Demonstrated ANN’s accuracy in predicting soil cohesion and friction angle using soil index properties	Limited dataset (200 samples); model performance dependent on input combinations
[123]	Review of ANN modeling and application issues	Outlined ANN architectures: BPNN, RNN, PNN, and SOM. Input selection, training/testing, and data preprocessing highlighted practical examples in liquefaction, pile capacity, soil classification, and slope stability	Provides a methodological framework for applying ANNs in geotechnical engineering	Issues: network geometry selection, data division, overfitting, computational demands
[24]	Review of AI optimization techniques (ANN, Fuzzy Logic, GEP, ANFIS, GA, etc.) in geotechnical applications	AI methods shown to improve prediction of soil behavior, pile capacity, swelling potential, foundation settlement, liquefaction, and more	Comprehensive overview showing how AI enhances geotechnical modeling, sustainability, and precision	Review article—no new experimental validation; heavy reliance on secondary data
[130]	Bootstrapping DL-ANN with airborne EM and borehole data for saltwater investigation in the Mississippi River Valley (USA)	Developed resistivity-to-lithology and resistivity-to-concentration models, while DL-ANN estimated the total dissolved solute	Results indicate salinity upconing due to excessive pumping	Reliance on used water quality data for training and validating the DL-ANN model. Inherent uncertainties with the transformation of resistivity values to lithologies and chloride concentrations
[131]	Review of ANNs in soil science	ANNs prove to be effective in predicting soil properties (pH, organic carbon, clay content, permeability, compaction, shear strength); useful for soil classification, fertility assessment, erosion prediction, and moisture estimation	Highlights ANN potential for soil modeling, land-use planning, and precision agriculture; relevant to geotechnical soil behavior predictions	Review only; lacks detailed experimental validation; applications are mostly limited to the soil science context

Table 1. Summary of selected reviews and research on ANN with its variants in EGGE.

Authors	Methods used	Results	Relevance	Limitations
[128]	Time Delay Neural Networks (TDNNs) applied to cyclic swelling data from the powerhouse cavern (Iran)	TDNN successfully modeled cyclic swelling pressure with good accuracy, capturing time-dependent behavior	Demonstrates ANN capability in modeling cyclic swelling/shrinkage of weak rocks (mudrock), critical for underground structures	Dependent on site-specific data, generalizability to other rock formations is uncertain
[129]	ANN (MLP, RBF) and multivariate regression with multivariate non-linear regression (MLR, MNR) to predict shear strength parameters in part of Iran	MLP-ANN outperformed RBF-ANN; MLR outperformed nonlinear regression; ANN captured complex nonlinear soil behavior	Demonstrated ANN’s accuracy in predicting soil cohesion and friction angle using soil index properties	Limited dataset (200 samples); model performance dependent on input combinations
[123]	Review of ANN modeling and application issues	Outlined ANN architectures: BPNN, RNN, PNN, and SOM. Input selection, training/testing, and data preprocessing highlighted practical examples in liquefaction, pile capacity, soil classification, and slope stability	Provides a methodological framework for applying ANNs in geotechnical engineering	Issues: network geometry selection, data division, overfitting, computational demands
[24]	Review of AI optimization techniques (ANN, Fuzzy Logic, GEP, ANFIS, GA, etc.) in geotechnical applications	AI methods shown to improve prediction of soil behavior, pile capacity, swelling potential, foundation settlement, liquefaction, and more	Comprehensive overview showing how AI enhances geotechnical modeling, sustainability, and precision	Review article—no new experimental validation; heavy reliance on secondary data
[130]	Bootstrapping DL-ANN with airborne EM and borehole data for saltwater investigation in the Mississippi River Valley (USA)	Developed resistivity-to-lithology and resistivity-to-concentration models, while DL-ANN estimated the total dissolved solute	Results indicate salinity upconing due to excessive pumping	Reliance on used water quality data for training and validating the DL-ANN model. Inherent uncertainties with the transformation of resistivity values to lithologies and chloride concentrations
[131]	Review of ANNs in soil science	ANNs prove to be effective in predicting soil properties (pH, organic carbon, clay content, permeability, compaction, shear strength); useful for soil classification, fertility assessment, erosion prediction, and moisture estimation	Highlights ANN potential for soil modeling, land-use planning, and precision agriculture; relevant to geotechnical soil behavior predictions	Review only; lacks detailed experimental validation; applications are mostly limited to the soil science context

3.2.2. Convolutional Neural Networks (CNNs)

CNNs are a leading DL architecture inspired by the visual cortex and widely applied in EGGE for spatial feature learning, enabling automatic multiscale pattern extraction from raw geospatial, geophysical, and imagery data [117]. As shown in Figure 6a–b, a typical AI–EGGE CNN workflow includes data preprocessing, augmentation, and transfer learning using pretrained models such as AlexNet, GoogLeNet, Inception, ResNet, U-Net, and DenseNet to enhance feature extraction and reduce training time and data needs. A standard CNN comprises convolutional and pooling layers with activation functions (e.g., ReLU), followed by either a flatten layer or global average pooling (GAP) before dense layers and a Softmax classifier, with Flatten adopted in Figure 6b. Flatten increases trainable parameters, whereas GAP reduces complexity and mitigates overfitting—beneficial in EGGE where labeled data are limited and heterogeneous [10,13].

CNNs outperform traditional ML methods based on handcrafted features by enabling end-to-end, noise-resilient representation learning [104,116]. Performance has been further enhanced through variants such as Visual Geometry Group (VGG), ResNet, and DenseNet for deep residual feature extraction [99,132]; Inception/GoogLeNet and EfficientNet for multiscale and computation-efficient learning [133]; MobileNet for lightweight deployment [106]; 3D-CNNs for volumetric subsurface data [134]; and U-Net, fully convolutional network (FCN), and SegNet for pixel-level segmentation [19,110]. Model evaluation typically uses accuracy, precision, recall, F1-score, and confusion matrices [13], while segmentation and generative CNNs additionally employ Intersection over Union (IoU), dice coefficient, peak signal-to-noise ratio, and Structural Similarity Index Measure (SSIM) [40,135]. High annotation needs, compute cost, and low interpretability are driving progress in transfer learning, few-shot learning, lightweight CNNs, and XAI/physics-informed CNNs [134,136].

CNNs show strong versatility across EGGE for material characterization, geospatial mapping, hazard assessment, and infrastructure monitoring. They routinely achieve >95–99% accuracy in soil and rock classification, texture analysis, and mineral discrimination using core, thin-section, UAV, and hyperspectral imagery [137,138]. Hemdan and Al-Atroush [127] demonstrated that CNN-based feature extraction combined with SVM improves soil-image recognition accuracy while reducing training cost. U-Net, SegNet, and FCN enable high-resolution segmentation of lithological boundaries and soil horizons, supporting digital soil mapping, fertility and pH assessment, and large-area soil property prediction [139,140]. In remote sensing, CNNs applied to ASTER, Landsat, Sentinel, and UAV data enhance lithological mapping, alteration detection, and landslide susceptibility analysis for early warning and resilience planning [111,141]. CNNs also deliver automated, high-accuracy crack and defect detection in tunnels, retaining walls, pavements, and embankments, outperforming manual inspection and traditional computer vision [142,143]. Hybrid CNN–LSTM/RNN models fuse spatial and temporal learning for slope-movement forecasting, deformation monitoring, and tunnel boring machine (TBM) ground-type recognition, with ResNet-18 and GoogLeNet reporting >96% accuracy for real-time TBM operations [54,133]. Compared with handcrafted image-processing techniques, CNNs provide superior multilevel feature learning and robust end-to-end classification; for instance, ResNet50 and VGG16 exceeded 98% accuracy in soil-aggregate classification [137], and 3D-CNNs delivered very high accuracy for hyperspectral soil imaging [144].

Despite the rapid progress, key gaps persist: cross-site generalization, label imbalance, and domain-shift sensitivity limit deployment, particularly in data-scarce or region-specific studies [13,145]. Emerging solutions include multimodal CNNs integrating geophysical, remote-sensing, and geotechnical imagery; CNN–TF hybrids for long-range spatial dependency learning; and physics-informed CNNs that embed geoscience constraints to suppress non-physical outputs [41,79]. Research is also advancing toward self-supervised and few-shot learning, synthetic data augmentation, and explainable CNNs to reduce labeling burden, improve robustness, and accelerate engineering adoption.

3.2.3. Recurrent Neural Networks (RNNs)

RNNs are DL architectures designed for sequential data, maintaining a hidden state that evolves over time to learn temporal dependencies between observations, an advantage over feedforward and CNN models, which process inputs independently or within finite receptive fields [117,144]. As shown in Figure 7 and expressed in Equations (1) and (2), each input $x_t$ is combined with the previous hidden state $h_{t-1}$ to produce an updated state $h_t$ [146], enabling RNNs to retain temporal memory essential for modeling evolving subsurface and geo-infrastructure behavior. This capability has driven their adoption in EGGE applications involving time-varying responses, including time-lapse geophysics, seismic site-response monitoring, pore-pressure and settlement prediction, slope-stability early warning from sensor streams, and tunnel and embankment health monitoring [147]. Practical deployment requires sequence normalization, handling irregular sampling, and managing concept drift in long-term monitoring [148,149]. To overcome vanishing and exploding gradients, gated variants such as LSTMs and Gated Recurrent Units (GRUs) enable long-range dependency learning and have become the dominant RNN models for sequential geotechnical data [73,150]. Despite their strengths, RNNs struggle with long-sequence cost, noise sensitivity, data drift, limited interpretability, and weak cross-site generalization. These limitations are driving advances such as hybrid CNN–RNNs, multimodal and self-supervised sequence models, and physics-informed RNNs for more robust engineering use. Their performance is typically assessed using time-series metrics (MSE, RMSE, MAE, MAPE) and correlation-based indices.

$$h_t=f\left(W_{xh}{x}_t+W_{hh}{h}_{t-1}+b\right)$$

(1)

$$o_t=g\left(W_{ho}{h}_t+c\right)$$

(2)

3.2.4. Deep RNN (DRNNs)

DRNNs extend conventional RNNs by stacking multiple recurrent layers to strengthen temporal feature learning and improve predictive accuracy [24,89,114]. In a typical DRNN architecture, inputs pass through successive RNN/LSTM/GRU layers, with hidden states propagated across time, enabling hierarchical learning of temporal patterns: lower layers capture short-term dynamics, while upper layers extract long-range and more abstract dependencies [149]. This depth advantage makes DRNNs well-suited for environmental and geotechnical problems involving nonlinear, multivariate time-series behavior. Applications include modeling rainfall–pore-pressure–suction interactions for slope-stability assessment, forecasting ground settlement under cyclic or construction loading, and fusing multi-sensor data for condition monitoring of tunnels, dams, and embankments [149,151]. By capturing both immediate and long-term dependencies in evolving geotechnical signals, DRNNs provide a more reliable basis for time-series prediction, early-warning analytics, and risk-informed geotechnical management.

3.2.5. RNN–Autoencoder (RNN–AE) and Bidirectional RNN (BiRNN)

In a typical RNN–Autoencoder (RNN–AE), the encoder compresses multivariate time-series into a latent vector that captures short- and long-term temporal dependencies, and the decoder reconstructs or forecasts sequences from this latent state [24,152]. RNN–AEs integrate the sequential learning ability of LSTM/GRU-based recurrent networks with the feature-learning and dimensionality-reduction strengths of autoencoders, enabling unsupervised representation learning from time-series data [149,153]. For example, the architecture described by Yu et al. [154] and Santoso et al. [153] stacks LSTM/GRU layers for encoding, applies a “RepeatVector” to expand the latent state to the forecast horizon, and employs a mirrored LSTM/GRU decoder with a “TimeDistributed” dense head for output generation. Model capacity is governed by hyperparameters N₁, N₂, and N₃ (units per layer), while dropout and layer normalization enhance training stability and generalization. Training is conducted using a composite loss that combines reconstruction mean squared error (MSE) with multi-step forecasting losses (MSE/MAE), enabling denoising, gap-filling, anomaly detection through reconstruction error, and multi-step nowcasting.

The other variant, Bidirectional RNNs (BiRNNs), processes sequences in both temporal directions to enhance prediction accuracy [149], by employing forward and backward hidden states that capture past and future inputs concurrently, with the outputs merged in a shared layer [114]. This dual perspective alleviates vanishing-gradient limitations and improves context awareness, making BiRNNs particularly effective for sequential learning tasks [133]. These capabilities make RNN–AEs and BiRNNs well-suited to EGGE applications involving noisy, irregular, and multivariate monitoring data, including time-lapse ERT/IP, pore-pressure and settlement logs, rainfall–infiltration–pore-pressure responses, geophysical monitoring, infrastructure-health sensing, and early-warning prediction in geophysical–geotechnical systems.

3.2.6. Gated Recurrent Unit (GRU)

GRUs are an advanced RNN variant for efficient time-series modeling, using gating mechanisms to retain relevant historical information while mitigating vanishing-gradient issues [114]. A GRU cell uses two gates—update and reset—to regulate information flow: the update gate controls how much past state is retained, while the reset gate determines how new input is combined with previous memory. Compared with LSTMs, which use three gates, GRUs have a simpler architecture with fewer parameters, enabling faster training and lower computational cost with minimal accuracy loss [155]. This parsimony makes GRUs suitable for real-time or resource-constrained EGGE workflows involving sensor-based monitoring, dynamic responses, and short- to medium-term forecasting. However, reduced gating complexity can limit long-term dependency modeling, where LSTMs often perform better. Overall, GRUs offer an effective balance of accuracy and efficiency, with the GRU–LSTM choice depending on length, data characteristics, and computational constraints.

3.2.7. Long Short-Term Memory (LSTM)

LSTM networks are an advanced RNN architecture designed to capture long-range temporal dependencies and overcome vanishing-gradient limitations in standard RNNs [112,118,156]. It addresses the problem of the backpropagated error either blowing up or decaying exponentially for long time lags in conventional RNNs. An LSTM cell maintains an internal cell state $c_t$ regulated by three gating mechanisms—input $\left(i_t\right)$, forget $\left(f_t\right)$, and output $\left(o_t\right)$ gates—which control how information is written, retained, and retrieved over time [117] (Figure 8). This gated structure stabilizes the cell state and enables selective preservation of relevant patterns, allowing LSTMs to model long-term sequential behavior more effectively than traditional RNNs and earlier models such as Hidden Markov Models [157,158]. LSTMs are particularly valuable for dynamic processes, e.g., rainfall–infiltration–suction cycles, pore-pressure and settlement evolution, structural health trends, etc., where retaining long temporal context improves forecasting accuracy [114]. However, under irregular or non-uniform time steps, simpler RNN variants may occasionally outperform LSTMs [133]. As shown in Figure 8, an LSTM cell receives the current input $x_t$ and previous hidden state $h_{t-1}$, which pass through the forget, input, and output sigmoid gates, as well as a candidate state generated via a $tanh$activation. The cell state is updated as (Equation (3)):

$$c_t=f_t\odot c_{t-1}+i_t\odot {\tilde{c}}_t,$$

(3)

while the hidden output is computed as (Equation (4)):

$$h_t=o_t\odot \mathrm{t}\mathrm{a}\mathrm{n}\mathrm{h}\left(c_t\right)$$

(4)

The gated memory pathway preserves gradients across long sequences, enabling stable training and reliable multi-step forecasting in noisy, nonlinear EGGE time-series data. Building on this capability, Bidirectional LSTM (BiLSTM) and Bidirectional GRU (BiGRU) extend the architecture by processing sequences in both forward and backward directions, allowing the model to learn from past and future context simultaneously [89]. At each time step, BiRNNs generate forward and backward hidden states that are fused to produce a context-enriched representation, particularly advantageous when complete sequences are available prior to inference, and when delayed outcomes depend on earlier and later events within the time window [133,154]. In EGGE, BiLSTM/BiGRU models have demonstrated superior performance compared to unidirectional variants by capturing fuller temporal dependencies in geotechnical signals, improving forecasting accuracy for slope-failure precursors, settlement evolution, seismic site response, TBM-induced vibrations, and structural-health trends in buried and surface infrastructure [133,149,159].

3.2.8. Generative Adversarial Networks (GANs)

GANs, introduced by Goodfellow et al. [160], are a key DL architecture for generative modeling. A GAN trains a generator $G$ to produce synthetic samples and a discriminator $D$ to distinguish them from real data through an adversarial minimax process [117], Figure 9a. Originally for unsupervised learning, GANs now support semi-supervised, supervised, and reinforcement learning, enabling data synthesis, super-resolution, domain translation, and anomaly detection [161,162]. Several variants enhance stability and performance: Deep Convolutional GAN (DCGAN) improves image quality using convolutional blocks; Wasserstein GAN (WGAN) stabilizes training and reduces mode collapse; Auxiliary Classifier GAN (ACGAN) enables label-conditioned generation; Variational Autoencoder GAN (VAE-GAN) fuses latent regularization with adversarial learning for sharper samples; Conditional GAN (cGAN) incorporates auxiliary data (e.g., lithology); CycleGAN performs unpaired image-to-image translation; and PatchGAN evaluates local patches to enhance fine-scale texture and spatial detail [30,161].

In EGGE, GANs help overcome data scarcity, improve model generalization, and enhance subsurface characterization. Key applications (Table 2) include synthetic augmentation of soil/rock datasets, reconstruction of incomplete borehole, seismic, and monitoring time series, pore-scale microstructure generation, and super-resolution of seismic and GPR imagery. Conditional and cycle-based GANs also enable cross-domain mapping (e.g., geophysical-to-lithology translation) [163]. Challenges persist—training instability, mode collapse, computational cost, and validation—but GANs remain promising for data enrichment, geo-imaging enhancement, and inversion and simulation support [30,164]. To address some of these limitations, Yan et al. [165] integrate CycleGAN-based cross-modal enhancement with hierarchical feature fusion to improve multi-sensor image integration (e.g., Vis–SAR) (Figure 9b), marking a shift toward reducing modality gaps and improving fusion consistency for high-fidelity geo-imaging.

Figure 9. (a) Schematic diagram of the GAN (adapted from [117]. (b) Typical advanced GAN-based cross-modal fusion framework integrating CycleGAN-driven cross-modal enhancement with hierarchical deep feature fusion (adapted from [165]). The model learns bidirectional mappings between visible (Vis) and SAR modalities via generators $G_{V\to S}$ and $G_{S\to V}$, with adversarial feedback from modality-specific discriminators.

Figure 9. (a) Schematic diagram of the GAN (adapted from [117]. (b) Typical advanced GAN-based cross-modal fusion framework integrating CycleGAN-driven cross-modal enhancement with hierarchical deep feature fusion (adapted from [165]). The model learns bidirectional mappings between visible (Vis) and SAR modalities via generators $G_{V\to S}$ and $G_{S\to V}$, with adversarial feedback from modality-specific discriminators.

Table 2. Summary of selected studies on GAN with its variants and their relevance in EGGE.

Author	Methods used	Results	Relevance	Limitations
[160]	Original GAN framework using generator + discriminator (MLP)	Demonstrated competitive sample generation on MNIST, CIFAR-10, and TFD datasets	Foundational to geotechnical applications (later adopted for soil/rock modeling, subsurface imaging)	Training instability, lack of explicit probability density, and mode collapse
[166]	GANs, cGANs, WGANs; applied to CIFAR-10 and medical images	Augmented datasets improved accuracy (e.g., CIFAR-10: +7.3%; medical imaging: +6.7%); significant FID/IS improvements	Framework directly transferable to geotechnical datasets (soil/rock images, seismic/GPR data)	Training instability, quality control of synthetic data, and high compute costs
[161]	Variants: DCGAN, WGAN, ACGAN, VAE-GAN; review of structural and loss-function improvements	Summarized enhancements that stabilize training & improve diagnostic accuracy	Shows potential of GANs in generating synthetic geotechnical signals (e.g., vibration data for fault diagnosis)	Training instability, domain transferability issues
[167]	Image-to-Image GAN (Pix2Pix) trained on synthetic RF data representing soil layers (sand, silt, clay); Soil behavior type index (Ic) used as input	Achieved mean absolute error (MAE) of 0.039 vs. 0.096 for nearest-neighbor interpolation; accurate for Ic < 3; demonstrated the feasibility of GANs for 2D soil schematization	Provides a novel AI-based approach for subsurface schematization, outperforming traditional interpolation; supports efficient soil classification and modeling with limited data	Based solely on synthetic data; performance biased toward datasets dominated by Ic < 3; requires balanced datasets; validation with real field data still needed
[163]	Multi-scale GAN (MS-GAN) for 3D geological modeling	Generated multiple 3D realizations capturing stratigraphy with quantified uncertainty	Highly relevant for site characterization, 3D subsurface modeling in geotechnics	Computationally intensive; depends on the quality of the training image; irregular boreholes require additional processing
[30]	SchemaGAN: cGAN with U-Net generator & PatchGAN discriminator; trained on 24,000 synthetic cross-sections and CPT-like data	Outperformed interpolation methods (MAE ≈ 0.039 vs. 0.096 for nearest-neighbor; better layer boundaries, anisotropy, and complex geometries); validated through blind expert survey and Dutch field case studies	Demonstrated robust, realistic subsurface schematization from sparse CPT data; scalable tool for geotechnical site characterization and digital twin applications	High computational training cost (95h on supercomputer); still reliant on synthetic training data; limited to 2D (needs extension to 3D); may struggle if field conditions deviate from training database

Table 2. Summary of selected studies on GAN with its variants and their relevance in EGGE.

Author	Methods used	Results	Relevance	Limitations
[160]	Original GAN framework using generator + discriminator (MLP)	Demonstrated competitive sample generation on MNIST, CIFAR-10, and TFD datasets	Foundational to geotechnical applications (later adopted for soil/rock modeling, subsurface imaging)	Training instability, lack of explicit probability density, and mode collapse
[166]	GANs, cGANs, WGANs; applied to CIFAR-10 and medical images	Augmented datasets improved accuracy (e.g., CIFAR-10: +7.3%; medical imaging: +6.7%); significant FID/IS improvements	Framework directly transferable to geotechnical datasets (soil/rock images, seismic/GPR data)	Training instability, quality control of synthetic data, and high compute costs
[161]	Variants: DCGAN, WGAN, ACGAN, VAE-GAN; review of structural and loss-function improvements	Summarized enhancements that stabilize training & improve diagnostic accuracy	Shows potential of GANs in generating synthetic geotechnical signals (e.g., vibration data for fault diagnosis)	Training instability, domain transferability issues
[167]	Image-to-Image GAN (Pix2Pix) trained on synthetic RF data representing soil layers (sand, silt, clay); Soil behavior type index (Ic) used as input	Achieved mean absolute error (MAE) of 0.039 vs. 0.096 for nearest-neighbor interpolation; accurate for Ic < 3; demonstrated the feasibility of GANs for 2D soil schematization	Provides a novel AI-based approach for subsurface schematization, outperforming traditional interpolation; supports efficient soil classification and modeling with limited data	Based solely on synthetic data; performance biased toward datasets dominated by Ic < 3; requires balanced datasets; validation with real field data still needed
[163]	Multi-scale GAN (MS-GAN) for 3D geological modeling	Generated multiple 3D realizations capturing stratigraphy with quantified uncertainty	Highly relevant for site characterization, 3D subsurface modeling in geotechnics	Computationally intensive; depends on the quality of the training image; irregular boreholes require additional processing
[30]	SchemaGAN: cGAN with U-Net generator & PatchGAN discriminator; trained on 24,000 synthetic cross-sections and CPT-like data	Outperformed interpolation methods (MAE ≈ 0.039 vs. 0.096 for nearest-neighbor; better layer boundaries, anisotropy, and complex geometries); validated through blind expert survey and Dutch field case studies	Demonstrated robust, realistic subsurface schematization from sparse CPT data; scalable tool for geotechnical site characterization and digital twin applications	High computational training cost (95h on supercomputer); still reliant on synthetic training data; limited to 2D (needs extension to 3D); may struggle if field conditions deviate from training database

3.3. Physics-Informed Neural Networks (PINNs) in AI–EGGE

PINNs are neural networks that solve partial differential equations (PDEs) by embedding physical laws or other governing constraints directly into the loss function, ensuring physics-consistent learning [168,169,170]. By converting these constraints into additional loss terms, PINNs integrate observational data with PDE residuals, initial, and boundary conditions to guide training, thereby enhancing generalization, reducing dependence on large labeled datasets, and improving interpretability relative to purely data-driven DL [23,171]. This capability is particularly relevant to EGGE, where subsurface behavior is governed by PDE-based geophysical and geotechnical laws [59,172]. Within the AI–EGGE paradigm, PINNs mark a shift from conventional pattern-recognition DL toward physics-informed learning, enabling models to retain mechanistic fidelity while leveraging data-driven flexibility [39]. Figure 10 illustrates a multi-receptive-field PINN (MRF–PINN) [171]. The initialized input field ($u_0$) is processed through parallel encoder–decoder branches with different receptive fields (${RF}_1-{RF}_6$), enabling multi-scale feature extraction. Each branch produces intermediate feature maps ($u_1-u_6$) and corresponding predictions ($p_1-p_6$), which are recombined via feature-map aggregation and linear superposition to form the MRF–PINN prediction field. Training is guided by a composite loss function, where weighting coefficients ${\mathsf{\lambda }}_{data}$, ${\mathsf{\lambda }}_{BC}$, ${\mathsf{\lambda }}_{PDEs}$, balance data misfit ($L_{data}$), boundary-condition constraints ($L_{BC}$), and physics-based PDE residuals ($L_{PDEs}$). This design supports stable learning of multi-scale, heterogeneous, and coupled subsurface processes characteristic of EGGE systems, e.g., [11,173].

PINNs have shown strong capability for forward and inverse modeling in EGGE, with applications in soil/rock consolidation [174], groundwater [59], seismic wave propagation [38], subsurface stress distribution [22], tunneling [172], and pile–soil interaction [175]. They deliver physically credible predictions under sparse or imperfect data and achieve accuracy comparable to analytical and numerical solutions, with growing value for parameter inference [138,176]. Ito et al. [177] estimated the coefficient of consolidation for Terzaghi’s 1D consolidation and unsaturated hydraulic properties directly from laboratory data. However, widespread adoption remains limited by reliance on synthetic data, reduced robustness under field noise and heterogeneity, training instability, slow convergence, loss-term sensitivity, and difficulty modeling highly nonlinear or coupled hydro-geomechanical processes [22,178,179]. Emerging advances, such as hybrid architectures (e.g., MRF-PINNs, operator-learning PINNs), adaptive loss balancing, integrated field–lab data training, uncertainty- and XAI-enhanced PINNs, and PINN–numerical solver coupling [171,180], are accelerating the shift toward trustworthy, physics-informed, and field-deployable AI for EGGE, laying the foundation for neural-operator models and real-time digital-twin subsurface systems. For example, Figure 11, Arif et al. [10], integrates an optimum GBM classifier with a PINN for circular slope-stability diagnosis: the GBM is trained on a 221-case dataset (inputs: $H/h$, slope angle $\beta$, cohesion $c$, unit weight $\gamma$, friction angle $\phi$, and pore-water pressure $r_u$) with 5-fold cross validation (CV), while the PINN enforces physics constraints to regularize learning. This hybrid achieved the strongest F1/κ–AUC performance and outperformed ANN, SVM, and RF baselines, demonstrating how physics guidance enhances generalization under limited data.

4. Emerging Trends in AI–EGGE

The increasing complexity and nonlinearity of subsurface systems now require analytical frameworks that can integrate diverse geophysical, geotechnical, remote-sensing, and environmental datasets into unified, physically consistent models. This need is driving a shift in AI–EGGE toward physics-informed, multimodal, explainable, and increasingly autonomous intelligence capable of capturing coupled subsurface processes and supporting real-time decision-making [181,182]. This shift is being enabled by the integration of ML/DL with physics-based modeling, multimodal data fusion, UQ, and edge-intelligent sensing, aimed at improving model reliability, transparency, and operational deployment [183,184]. At the core of this evolution is multimodal fusion, which integrates heterogeneous datasets to reconcile differences in sensitivity, resolution, and noise [4,185], thereby reducing interpretational ambiguity and strengthening predictive robustness for environmental and geotechnical decision-making.

4.1. Explainable AI (XAI): Concept and Model Framework

As AI systems become embedded in geophysical–geotechnical workflows, interpretability is essential for scientific credibility, regulatory acceptance, and geoenvironmental and geoengineering deployment. In XAI, the terms white-box, gray-box, and black-box denote increasing opacity in a model’s internal logic and decision-making process [186], influencing model selection in EGGE applications. White-box models are inherently interpretable but often less accurate; gray-box models balance interpretability and performance; and black-box models typically achieve higher accuracy at the expense of transparency. XAI promotes transparency by linking model predictions to physical properties, uncertainty sources, and domain knowledge, converting opaque outputs into defensible analytical insights. It aims to make AI results understandable and trustworthy for both domain specialists and system developers, supporting error diagnosis, model refinement, and responsible use in high-consequence contexts [9,79,186]. XAI methods fall into two groups: intrinsically interpretable models designed for transparency (e.g., EBM, NAM, CBM) and post-hoc techniques that explain complex models (e.g., SHAP, LIME, IG, Grad-CAM, ALE) [187,188]. Beyond transparency, XAI supports scientific consistency by ensuring that learned representations align with physical principles and empirical evidence critical in EGGE, where decisions must withstand scrutiny under partial observability and high impact [189]. Table 3 summarizes the key XAI and uncertainty evaluation methods essential for interpreting multimodal models and assessing their predictive reliability.

4.1.1. Post-Hoc Methods

Post-hoc methods explain trained models at local or global levels. Key feature-attribution tools include Integrated Gradients and its refinement, Guided IG, which reduces attribution noise, and Shapley-value methods, with dependence-aware variants preferred when predictors are correlated [188,189]. For global effects, ALE is generally favored over Partial Dependence Plots (PDPs) as it avoids extrapolation under correlation, with recent work improving its efficiency in high-dimensional settings. These methods support factor-effect interpretation in geospatial workflows involving terrain, soil, or geophysical attributes [56]. However, post-hoc explanations may be unstable, inconsistent, or unfaithful to model logic; thus, quantitative tests of faithfulness and stability are recommended over visual inspection alone [81,190]. Demonstrated value in EGGE includes landslide susceptibility modeling, where SHAP-based analyses reveal the influence of slope, precipitation, and lithology, with similar usage in hazard mapping, subsoil classification, and monitoring [80,191].

4.1.2. Intrinsic (Ante-Hoc) Methods

Intrinsic approaches embed interpretability in the model form and are advantageous for validation in safety-critical engineering contexts. Modern additive models include NAMs, which retain GAM-like clarity with neural flexibility, and EBMs, which provide transparent, shape-constrained feature effects with competitive accuracy [180,187]. CBMs introduce human-auditable intermediate concepts to enable inspection and correction at the concept layer [192]. A related direction leverages physics-informed ML (PIML) or physics-guided ML, embedding physical laws, constraints, or forward operators into the architecture or loss to yield predictions interpretable through constraint satisfaction and residuals—useful for inversion and property mapping in EGGE [101,169,180].

4.1.3. Integrated Applications

Applications in EGGE demonstrate the value of XAI for improving model credibility, interpretability, and deployment. For instance, Wang et al. [80] employed an XAI–DL framework to derive global and local insights into hydro-morphological processes (HMP) across China to aid in hazard and risk assessment, achieving ten-fold cross-validated AUC scores of 0.83–0.86. The SHAP-based interpretations revealed that spatially varying feature contributions to HMP predictions can diagnose model behavior, validate physically meaningful drivers, and support regional-scale environmental analysis. In landslide susceptibility modeling, e.g., [80,188], interpretable neural networks and SHAP-augmented ensembles similarly expose dominant predictors and interactions (e.g., slope–precipitation coupling, lithology, aspect), improving user trust and guiding model refinement. In geotechnical property prediction and soil classification, SHAP, ALE, and partial-dependence diagnostics clarify how CPT-derived indices and soil parameters influence outputs [9,16]. Similarly, Degen et al. [180] show that architectures that embed domain constraints into feature learning enhance classification performance while preserving the interpretability of the physical cues used. Overall, XAI strengthens due diligence in EGGE, supporting sensitivity and scenario analysis, defensible decision-making, and model-risk management across heterogeneous datasets [189].

The practical guidance XAI interpretations include:

• Use intrinsic models (EBM, NAM, CBM) where interpretability is essential; apply post-hoc methods mainly for auditing, debugging, and communicating complex models, and validate explanations with faithfulness and stability checks.
• In correlated multimodal settings, favor ALE and dependence-aware SHAP over naïve PDP/SHAP pipelines.
• Where physical laws are well established, adopt physics-informed hybrids to couple accuracy with mechanistic interpretability and diagnosable residuals.
• Quantify uncertainty in explanations (e.g., confidence intervals for feature effects or attribution variability across perturbations) to avoid overconfident interpretation in high-stakes EGGE decisions.
• Communicate explanations at the appropriate level of abstraction for the audience (engineers, regulators, field practitioners), emphasizing actionable insights rather than raw attribution maps.

4.2. Multimodal Fusion: Concept, Hierarchy, Strategies, and Methodological Approaches

Multimodal fusion in AI–EGGE refers to the systematic integration of heterogeneous datasets to achieve a more complete, reliable, and physically consistent representation of subsurface and engineered environments [4,193]. The conceptual hierarchy of multimodal fusion spans three integration levels: low-, mid-, and high-level fusion. Low-level (data-level) fusion combines raw or minimally processed datasets to preserve intrinsic spatial and physical relationships; mid-level (feature-level) fusion generates shared latent representations by combining engineered or learned features to improve discrimination of lithological boundaries, hydrostratigraphic units, geomechanical heterogeneities, and deformation precursors; and high-level (decision-level) fusion aggregates outputs from independently trained models through ensemble learning, Bayesian model averaging, and probabilistic late-fusion schemes to strengthen robustness, interpretability, and out-of-distribution generalization [4,21,182,183,194]. This conceptual multimodal fusion framework is illustrated in Figure 12, summarizing the flow from data acquisition and preprocessing through fusion levels and AI modeling to interpretability, uncertainty assessment, and end-use applications.

Relevant data sources include geophysical surveys (resistivity, seismic, GPR, magnetic, gravity), geotechnical investigations (borehole logs, SPT-N, CPT-qc, laboratory tests), and remote-sensing and environmental observations such as UAV imagery, light detection and ranging (LiDAR), synthetic aperture radar (SAR), interferometric SAR (InSAR), and soil chemistry. Each modality contributes distinct yet complementary strengths: geophysical methods provide broad spatial coverage but often suffer from non-uniqueness and depth-resolution limitations, whereas geotechnical data offer high-fidelity ground truth but are spatially sparse. Their fusion reduces interpretational ambiguity, tightens model uncertainty, and enhances the effective spatial resolution and structural interpretability of subsurface models for geoenvironmental and geoengineering decision-making [193,195,196]. Table 4 provides a complementary overview focusing specifically on the hierarchy of fusion levels, representative techniques, and computational enablers adopted in EGGE.

Methodologically, multimodal integration in EGGE is enabled through five overarching strategy classes: physics-governed fusion (e.g., joint and cooperative inversion) that embeds physical constraints; data-driven fusion using ML/DL architectures; hybrid physics–AI fusion that couples domain knowledge with learning-based models; cross-modal alignment strategies that learn shared latent spaces for correlated modalities; and probabilistic and uncertainty-aware fusion to quantify confidence and support risk-informed decision-making [45,182]. Recent advances further incorporate AI-driven approaches, including physics-informed ML (PIML/PINNs) for physics-consistent multimodal learning, transformer-based and contrastive cross-modal fusion for spatial-temporal representation alignment, diffusion and GAN models to infer or reconstruct missing modalities, and GNN- and spatiotemporal fusion frameworks for continuous deformation assessment, infrastructure monitoring, and hazard early warning [165,185,194]. Deep attention-enhanced fusion architectures are also emerging, exemplified by Yan et al. [165], who integrate multi-layer deep feature extraction, hierarchical feature fusion with attention, and a Layered Strictly Nested Framework (LSNF) to achieve robust SAR–visible image fusion with preserved structural integrity and high-frequency detail. Recent evidence shows that integrating borehole and laboratory data with ground and airborne geophysics sharpens hydrostratigraphic and deformation models and strengthens prediction in EGGE [195,196]. Physics-consistent joint inversion further improves structural coherence in integrated subsurface interpretation [193,198]. Remote-sensing fusion (optical/SAR/LiDAR) adds terrain, deformation, and environmental constraints at scale, supporting hazard-prone and urban-infrastructure applications [4,64,199].

5. Applications and Case Studies of AI–EGGE

5.1. Applications

The practical deployment of AI–EGGE has advanced from isolated algorithmic trials to robust, field-validated decision-support systems. Beyond conventional supervised and unsupervised ML, recent developments integrate DL architectures, physics-informed and mechanistic modeling, multimodal and cross-domain data fusion, and XAI to enhance trust, interpretability, and engineering uptake. These advances now enable high-resolution subsurface characterization, soil and lithological classification, geotechnical parameter prediction, groundwater assessment, geohazard forecasting, underground infrastructure monitoring, and real-time sensing for resilient engineering design. This section presents application cases demonstrating how AI–EGGE is transforming data acquisition, interpretation, and risk-informed decision-making across environmental and engineering contexts.

5.1.1. Site Characterization and Subsurface Profiling

AI–EGGE strengthens site characterization by improving assessment of soil–rock interactions and subsurface conditions, reducing trial-and-error investigations. It enables reliable prediction of groundwater levels, soil texture, particle-size fractions, organic carbon, pore-water pressure, and soil movement, lowering exploration costs and improving design efficiency. Accurate subsurface characterization is critical because physical, mechanical, and hydraulic properties govern structural performance [6,14,87]. Compaction, shear strength, and permeability control foundation support, while interactions—such as water content, viscosity, and earth pressure in clays—affect strength and deformation. Data-driven and statistical methods enhance characterization by revealing patterns often missed by conventional approaches.

ML models have been widely applied to estimate soil hydraulic properties [100,200], soil organic carbon and matter [201], and soil pore structures [53,77]. Notable examples include ANNs and the Group Method of Data Handling (GMDH), a self-organizing polynomial-based ML method, which have provided deeper insight into site properties [75,108,202]. Licznar and Nearing [203] applied ANNs to predict hydraulic behavior, and Mishra et al. [204] used RF and SVM to estimate bulk density and cation exchange capacity at different depths. Extreme learning machine (ELM), RBF, modified GMDH, and M5 tree methods have proven effective for predicting pore-water pressure, groundwater table elevation, and soil-water retention curves (SWRC), which are central to hydrological processes and soil moisture dynamics, and agro-hydrology [205]. A wide range of soil attributes—from texture and organic matter to hydraulic and retention properties—serve as key inputs for ML-based SWRC estimation. ANN- and DNN-based models have also been used to predict soil temperature [42,206].

Despite advances, site-specific soil and rock heterogeneity still necessitate complementary laboratory testing. Consistently with this, Wang et al. [207] assessed RF, SVM, gradient-boosted regression trees, MLP, and least angle regression (LAR) for hyperspectral-based soil salinity prediction, with LAR showing superior stability and accuracy. Elaziz et al. [208] also applied XGBoost, RF, and GBM to 120 soil samples for predicting soil salinity in semi-arid regions, with a success rate between 0.99% to 1.0%. Zhang et al. [209] reported that XGBoost achieved the highest accuracy for total soil nitrogen prediction, RF performed best with SAR data, and GBM outperformed other models when using Landsat-8 and Sentinel-2 data compared to Sentinel-1. These highlight the value of integrating ML with multisource satellite data for precise soil nutrient mapping and improved geotechnical prediction, risk assessment, design optimization, and decision-making. Table 5 provides additional AI–EGGE studies on soil prediction and classification.

5.1.2. Landslides

Landslide research is grouped into susceptibility assessment, displacement prediction, and detection. Landslides may be shallow or deep-seated and commonly triggered by rainfall, earthquakes, typhoons, or deforestation [212]. Landslide susceptibility mapping (LSM) uses conditioning factors such as elevation, slope, lithology, soil type, aspect, and distance to roads, faults, and rivers, with steep slopes, weak soils, low shear strength, and river proximity increasing susceptibility [12,213]. Detection maps event location, type, boundary, volume, date, and soil/rock triggers, supporting regional “soft” risk assessment, whereas displacement prediction represents “hard” risk evaluation for specific slopes or assets [214,215]. Traditional detection relied on visual imagery and field surveys [216]. The Frequency Ratio method remains widely used to quantify links between conditioning factors and landslide occurrence, with correlation/multicollinearity checks applied to reduce redundant predictors [217].

AI-driven semi- and fully automated workflows now reduce manual effort and enhance detection and LSM efficiency, especially for pixel-based and object-based image analysis (OBIA) methods [110,213,215]. OBIA generally outperforms pixel-based methods in complex terrain by detecting object-level changes from multi-temporal imagery. For instance, Ghorbanzadeh et al. [110] employed a hybrid ResU-Net–OBIA workflow for landslide detection. In the study, Sentinel-2, NDVI, and slope layers are first segmented using multiresolution segmentation, followed by rule-based OBIA classification to generate object-level candidates. In parallel, ResU-Net performs pixel-wise segmentation to produce probability heatmaps, which are thresholded and fused with OBIA outputs. The hybrid ResU-Net–OBIA approach reduces spectral confusion, enhances boundary delineation, and suppresses false positives compared with standalone ResU-Net or OBIA, demonstrating the value of combining deep semantic segmentation with object-based geomorphological reasoning for landslide mapping.

In other studies, Liu et al. [218] demonstrated that the YOLOv7 model enhanced with a Squeeze-and-Excitation attention mechanism provided higher generalization capability and detected landslides more accurately with fewer missed events. CNNs and RNNs consistently outperform SVM, DT, and RF for LSM [219,220,221]. Althuwaynee et al. [222] found that a hybrid evidential-AHP (analytical hierarchy process) model outperformed logistic regression. Wang et al. [223] reported that an ensemble of SVM, ANN, and GBoost improved AUC by 0.11–0.35 over single models. Kim and Lee [206] compared CNNs with ML models (RF and bagging) and showed that CNNs consistently achieved higher LSM accuracy. Balogun et al. [102] used meta-heuristic algorithms—GWO, Bat Algorithm (BA), and Cuckoo Optimization Algorithm (COA)—to optimize SVM hyperparameters, improving LSM prediction accuracy. GRU, CNN, and hybrid models (CNN–AEIO [Age of Exploration-Inspired Optimizer], PSO–SVM, SVM–GWO) have also shown strong LSM performance in Lushan, Majiagou, northwest Sichuan, and Jiuxianping [76,106,119,224].

For time-series-based landslide modeling, DNNs, TF-based temporal models, and physics-informed hybrids have improved landslide displacement forecasting; however, models such as GEP, SVM, and RBF often perform poorly when external triggers (e.g., rainfall and reservoir water-level fluctuations) are excluded, limiting their long-term predictive ability [215,225]. GEP has shown superior performance to SVM and RBF in displacement prediction [85]. Comparative DL studies report that transformer models yield the highest accuracy, followed by LSTM, improved Elman networks, RNNs, and other DL architectures [136,158]. The improved Elman network, which incorporates a piecewise time-weighted gradient function, enhances the ability to learn both current and historical temporal dependencies [225].

5.1.3. Sinkholes

Sinkholes are vertical depressions formed by the collapse of overlying soil due to chemical dissolution of underlying karstic bedrock, particularly in carbonate and evaporite terrains [9,214,226]. They display distinct morphologies, such as depth, diameter, circumference, and area, that vary with formation environment and lithology (e.g., basement rock, plateau, and basin sinkholes. Efficient detection and analysis of sinkhole patterns are essential for disaster mitigation and sustainable development [19,227]. ML/DL applied to subsurface datasets provides a robust means of characterizing sinkhole morphology and distinguishing them from surrounding terrain. Integrated ML/DL models have leveraged diverse data sources, including aerial photographs, GPR, thermal imagery, digital elevation models (DEMs), satellite data, and RGB images [228,229,230,231]. In West-Central Florida, Muili and Babaie [232] showed that RF outperformed MLP, SVM, and KNN for sinkhole prediction, and revealed that shallow bedrock depth, land-use patterns, and NW–SE-trending faults were the dominant controls on sinkhole development. Table 6 further provides additional studies on AI applications for sinkhole detection using remote sensing data.

5.2. Case Studies

This section presents real-world cases showing how AI enhances subsurface characterization, hazard prediction, and engineering decisions in EGGE. Each case summarizes: (i) the problem context, (ii) the AI/ML/DL, hybrid, or physics-informed models applied, (iii) multimodal or multisensor data used, and (iv) the value-added gains over conventional methods. These cases reflect the AI–EGGE shift toward data-driven, physics-consistent, and multimodal intelligence for more accurate and decision-ready solutions.

5.2.1. Case 1: Hybrid CNN–ViTF for High-Fidelity, Robust, and Real-Time ERT Inversion

ERT inversion has long been constrained by smoothing artifacts, non-uniqueness, and high computational demand inherent to Gauss–Newton (GN)–based solvers. Yin et al. [41] introduced a Hybrid ViTF–based inversion framework, with spatial CNN blocks and TF self-attention, that transitions ERT imaging from iterative physics-based inversion to a direct, data-driven, image-to-image prediction paradigm with high structural fidelity and real-time performance. The workflow (Figure 13) integrates: (i) large-scale synthetic paired apparent–true resistivity datasets generated through finite-element forward modeling for 1–5 anomalous targets, enabling curriculum training to progressively learn structural complexity; (ii) a residual CNN-based colormap calibration module to predict resistivity ranges for field data where true values are unavailable, ensuring consistent pixel–resistivity mapping; and (iii) the ViTF inversion model that maps calibrated apparent resistivity images to true resistivity distributions (Figure 13a). The detailed ViTF architecture is further illustrated, combining convolutional blocks (Figure 13b) for localized anomaly feature extraction with TF self-attention maps observed apparent-resistivity pseudo-sections to true resistivity distributions for resolving and capturing long-range resistivity dependencies. This architecture overcomes the locality bias and scale sensitivity of CNN-only models and enables generalization across anomaly geometries, contrasts, and spatial configurations. A comparative performance summary of ViTF against CNN-AE, U-Net, Latent Diffusion Model (LDM), and GN inversion confirmed its superior accuracy, efficiency, boundary preservation, and robustness across all complexity levels.

The ViTF achieved SSIM up to 0.912, MSE as low as $1.12\times 10⁻³$, and the shortest training time of 706 s, outperforming CNN-AE and U-Net, which produced blurred anomaly boundaries and higher errors, and outperforming LDM, which showed instability beyond one-target scenarios. It maintained R² > 0.96 and RMSE < 260 Ωm across 1–5 targets, demonstrating stable high-fidelity inversion with sharp boundary retention. In Figure 14, the colormap calibration module (Panel 1) delivered R² > 0.98, MAE < 150 Ωm, and RMSE < 260 Ωm, confirming negligible deviation between calibrated and ground-truth colormaps, while inversion results remained statistically unchanged (Panel 2), validating field readiness. Against GN inversion, the ViTF (Panel 3) consistently produced lower pixel-level errors, sharper target geometries, and reduced artifact zones across all five complexity cases, whereas GN exhibited smearing and centroid misplacement of anomalies. Notably, ViTF achieved ~20 ms per inversion, compared to ≥ 5 s per iteration for GN, marking a transition toward real-time ERT imaging. In field validation at the U.S. DOE Hanford Site, ViTF successfully reconstructed key resistivity contrasts associated with stratified vadose-zone heterogeneity and contamination pathways, outperforming GN, which underestimated boundary sharpness and mispositioned anomaly interfaces. While current limitations include reliance on synthetic training data, future integration with PINNs, uncertainty-aware diffusion models, and multimodal joint inversion (e.g., ERT–seismic–GPR) will further enhance generalizability across lithological regimes. This TF-based framework exemplifies the AI-EGGE paradigm shift, where DL models evolve from post-processing aids to primary inversion engines capable of augmenting or replacing conventional physics-based solvers. The ViTF establishes a new benchmark for rapid, high-resolution, and operationally deployable ERT inversion, enabling real-time decision support for environmental, geotechnical, and subsurface hazard monitoring applications.

5.2.2. Case 2: Multimodal CNN–TF Fusion for Enhanced Urban Scene and Functional Mapping

Urban scene understanding is critical for land-use mapping, infrastructure planning, and exposure assessment in complex environments; however, remote sensing imagery (RSI) alone often struggles to discriminate morphologically similar urban classes. Su et al. [4] addressed this limitation through a multimodal CNN–Transformer fusion framework that integrates RSI with Points of Interest (POIs) and building footprint data, enabling semantically enriched urban functional mapping. The model employs a dual-branch architecture with attention-weighted fusion and multiscale feature extraction, while a transformer-based interaction module aligns cross-modal representations to enhance feature consistency and contextual awareness.

Experimental results on the Chengdu and Wuhan datasets demonstrate that multimodal integration improves overall accuracy by approximately 6–7% over RSI-only baselines, with consistent gains in Cohen’s Kappa, κ, and F1-score. Ablation analysis confirms that attention weighting, modality interaction, and multiscale feature extraction each contribute measurably to performance, while cross-city validation further indicates improved generalization relative to single-modality approaches. Comparative evaluations show that the framework achieves competitive or superior accuracy relative to recent multimodal architectures, with lower model complexity and improved interpretability. Interpretability analysis using Class Activation Maps reveals that multimodal fusion produces more spatially coherent and semantically meaningful attention patterns, particularly for complex functional classes such as commercial, public service, and high-density residential zones, where RSI-only models exhibit ambiguous or diffuse responses. The lightweight design (~6.23M parameters) supports scalable deployment for large-area urban monitoring; however, reliance on static datasets limits temporal adaptability, highlighting the need for integrating dynamic data sources such as mobility patterns and multi-temporal imagery. This case study demonstrates how multimodal fusion within the AI-EGGE paradigm integrates heterogeneous spatial data to improve classification reliability, interpretability, and generalization, strengthening urban analytics for infrastructure exposure assessment and climate–geohazard-informed planning.

6. Challenges and Limitations

Despite major advances, AI–EGGE faces persistent data and fusion-related barriers that limit reliable deployment. Data scarcity, uneven quality, and limited labeled samples—particularly in environmental geophysics, near-surface imaging, and hazard-related applications where ground truth is costly or uncertain—continue to constrain model performance [41,199]. Multimodal fusion must reconcile heterogeneity in spatial support, sampling density, sensor footprint, and noise, especially in shallow, heterogeneous environments, complicating co-registration, scaling, and correlation handling and potentially biasing fused estimates [1,4]. Class imbalance and spatial autocorrelation can inflate apparent accuracy and obscure generalization, while inconsistent metadata, weak documentation, and the lack of standardized multimodal datasets hinder reproducibility, benchmarking, and cross-site transferability [30].

Model robustness, physical consistency, and interpretability remain insufficient for engineering-grade decision-making. Many models trained on site-specific datasets degrade under domain shift, non-stationarity, or extreme triggers, with limited cross-site validation and no widely accepted benchmark datasets for fair comparison [4,100]. Purely data-driven models may violate physical constraints and fail outside the training envelope; although physics-informed, hybrid, and joint inversion approaches improve realism, they remain computationally intensive, often ill-posed, and susceptible to non-unique solutions without informative coupling or priors [115,176]. Modal fusion and deep fusion models raise additional interpretability and black-box concerns for engineering decisions, reinforcing the need for physics-informed and XAI-enhanced mechanisms to produce physically meaningful reasoning for spatiotemporal and multimodal systems—especially for environmental geophysics applications where shallow targets demand high sensitivity to noise, heterogeneity, and uncertainty [186,192]. UQ is rarely calibrated, limiting confidence and safety-margin assessment for risk-sensitive EGGE applications [165,182].

Operational, regulatory, and institutional barriers continue to slow the translation of AI–EGGE from research to practice. Developing, validating, and maintaining deep or physics-informed models requires specialized expertise, computing resources, and data-annotation capacity that many engineering and environmental agencies lack, while the absence of lightweight, real-time, edge-deployable tools limits early-warning, digital-twin, and on-site applications [17,134,194]. Data sharing restrictions, unclear responsibility for AI-assisted decisions, and limited long-term field pilots further reduce confidence and uptake [186]. In addition, widely relied-upon environmental and engineering standards—such as the European Standards for Structural and Geotechnical Design (Eurocode), the American Association of State Highway and Transportation Officials specifications, the Japanese Geotechnical Society Standards, the Australian Standards for geotechnical and structural design (including AS 1726 and AS 5100), and Environmental Impact Assessment (EIA) regulatory frameworks, currently provide no formal provisions for AI-assisted analyses, constraining regulatory acceptance in design, compliance, and risk management. Similar gaps exist across regional systems, including EIA regimes in Asia, Environmental Management Acts in Africa, and the Environmental Protection and Biodiversity Conservation Act in Australia, which also lack defined pathways for AI-enabled evaluation and decision support. Ethical, cybersecurity, and misuse risks—especially for automated or safety-critical environmental and infrastructure systems—remain unresolved, collectively limiting safe and regulated deployment of AI–EGGE.

7. Future Directions

Advancing AI–EGGE requires moving beyond data-driven prediction toward physics-grounded, uncertainty-aware, real-time, and decision-support AI suitable for field and engineering deployment. Priority lies in maturing physics-integrated learning—through PIML/PINNs, operator-learning networks, and theory-guided hybrids—to reduce non-uniqueness, improve generalization, and stabilize multimodal inversion by embedding governing equations, constitutive relations, and conservation laws directly into learning, even under sparse or noisy data [23,39,171]. Parallel advances in multimodal fusion are expected, with cross-modal attention, foundation-scale multimodal TFs, and GNNs enabling robust integration of geophysical, geotechnical, environmental, and remote-sensing data with differing spatial supports and noise characteristics [4,136,181]. These architectures encode spatial topology, physical neighborhoods, and causal interactions while preserving subsurface structural integrity [108]. When coupled with calibrated uncertainty quantification, interpretability, and physics-based regularization, these architectures can deliver trustworthy models for engineering design, hazard assessment, environmental monitoring, and subsurface digital twins [51,238].

A further frontier is real-time, autonomous, and adaptive geosensing. Edge–cloud computing, IoT-enabled instrumentation, UAV/robotic survey automation, and streaming geophysics will allow continuous assimilation of data for early warning, infrastructure monitoring, and dynamic model updating [53,149,239]. Active learning and adaptive survey designs—guided by model uncertainty—will optimize field campaigns by identifying where to drill, scan, or image to maximize information gain while minimizing cost [1,148]. Progress will also depend on building community-driven multimodal benchmarks, open datasets, and validation protocols tailored to EGGE tasks, along with regulatory and standards development to support safe engineering adoption. Ultimately, the integration of interpretable AI, physics-informed modeling, and real-time digital-twin ecosystems will position AI–EGGE as a core enabler of climate resilience, sustainable urban development, and risk-aware geotechnical and environmental engineering.

Strategic Roadmap for Advancing AI–EGGE:

6.

Physics-aligned and trustworthy intelligence: Unify physics-integrated AI (PIML, PINNs, neural operators) with calibrated UQ and interpretable/XAI frameworks to ensure physically consistent, reliable, and audit-ready predictions for engineering-grade deployment.

7.

Multimodal fusion and adaptive data ecosystems: Develop next-generation fusion using TFs, GNNs, and cross-modal learning to harmonize geophysical, geotechnical, environmental, and remote-sensing data, supported by adaptive, uncertainty-guided field acquisition that maximizes information efficiency.

8.

Autonomous, real-time, and digital-twin EGGE systems: Establish edge–cloud AI platforms, IoT sensing networks, UAV/robotic acquisition, and continuous monitoring to enable real-time hazard detection, early warning, and autonomous digital-twin subsurface systems for resilient infrastructure and environmental management.

9.

Standardization, benchmarking, and engineering integration: Create open multimodal benchmarks and validation protocols and co-develop practice-ready AI–EGGE workflows with industry and regulators—embedding safety, due diligence, and codes-of-practice alignment to accelerate formal adoption into engineering standards.

8. Conclusions

AI has catalyzed a decisive shift in EGGE from empirical and deterministic approaches toward data-rich, physics-consistent, and decision-oriented subsurface intelligence. This review synthesizes advances in ML/DL, physics-informed and theory-guided modeling, multimodal data fusion, uncertainty-aware and XAI, and intelligent sensing. Together, these developments enable more accurate and scalable solutions for site characterization, lithological and geotechnical profiling, joint hydrogeomechanical estimation, groundwater and hydro-environmental assessment, geohazard detection, and infrastructure condition monitoring. This trajectory reflects the maturation of AI–EGGE from exploratory algorithms to validated, field-deployable systems supporting early warning, risk-informed design, sustainable infrastructure, and climate-resilient environmental management. The convergence of geophysical, geotechnical, environmental, and data-centric domains now positions AI–EGGE as a core pillar of next-generation subsurface characterization and hazard mitigation.

Responsible, regulated, and engineering-grade deployment, however, remains a critical frontier. Enduring limitations, including data heterogeneity, cross-scale incompatibility, sparse ground truth, lack of standardized multimodal benchmarks, weak generalization across geological and climatic regimes, inadequate interpretability, and limited UQ, continue to constrain confidence, regulatory acceptance, and industry-wide adoption. The absence of formal validation pathways for AI-assisted analyses within environmental and engineering codes and standards underscores the need for auditable modeling frameworks and safety-critical oversight. Future progress must accelerate the integration of physics-based and uncertainty-aware architectures, cross-modal attention and graph-based fusion, digital-twin ecosystems with edge–cloud intelligence, and adaptive sensing strategies that reduce data redundancy while maximizing information value. Realizing the full potential of AI–EGGE requires coordinated efforts across research, practice, regulatory bodies, and international standard-setting organizations to embed rigor, transparency, ethics, and public trust, enabling its transition from a promising innovation to a blueprint for resilient, autonomous, and sustainability-aligned subsurface intelligence.