Next Article in Journal
A Hydrogen Energy Storage Configuration Method for Enhancing the Resilience of Distribution Networks Within Integrated Energy Systems
Previous Article in Journal
A Differential Planning Strategy for Distribution Network Resilience Enhancement Considering Decision Dependence Uncertainty
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Energies
Volume 18
Issue 23
10.3390/en18236354
energies-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Review

Artificial Intelligence for Underground Gas Storage Engineering: A Review with Bibliometric and Knowledge-Graph Insights

by
Jiasong Chen
1,
Guijiu Wang
1,
Xuefeng Bai
1,
Chong Duan
1,
Jun Lu
1,
Luokun Xiao
1,
Xinbo Ge
2,*,
Guimin Zhang
3 and
Jinlong Li
4
1
Pipechina Energy Storage Technology Co., Ltd., Shanghai 200011, China
2
College of Energy and Mining Engineering, Shandong University of Science and Technology, Qingdao 266590, China
3
School of Mechanics and Civil Engineering, China University of Mining and Technology, Xuzhou 221116, China
4
College of Civil Engineering and Architecture, Zhejiang University, Hangzhou 310058, China
*
Author to whom correspondence should be addressed.
Energies 2025, 18(23), 6354; https://doi.org/10.3390/en18236354
Submission received: 9 October 2025 / Revised: 2 November 2025 / Accepted: 1 December 2025 / Published: 3 December 2025
(This article belongs to the Section D: Energy Storage and Application)
Browse Figures
Versions Notes

Abstract

Underground gas storage (UGS), encompassing hydrogen, natural gas, and compressed air, is a cornerstone of large-scale energy transition strategies, offering seasonal balancing, security of supply, and integration with renewable energy systems. However, the complexity of geological conditions, multiphysics coupling, and operational uncertainties pose significant challenges for UGS design, monitoring, and optimization. Artificial intelligence (AI)—particularly machine learning and deep learning—has emerged as a powerful tool to overcome these challenges. This review systematically examines AI applications in underground storage types such as salt caverns, depleted hydrocarbon reservoirs, abandoned mines, and lined rock caverns using bibliometric and knowledge-graph analysis of 176 publications retrieved from the Web of Science Core Collection. The study revealed a rapid surge in AI-related research on UGS since 2017, with underground hydrogen storage emerging as the most dynamic and rapidly expanding research frontier. The results reveal six dominant research frontiers: (i) AI-assisted geological characterization and property prediction; (ii) physics-informed proxy modeling and multi-physics simulation; (iii) gas–rock–fluid interaction, wettability, and interfacial behavior prediction; (iv) injection-production process optimization; (v) intelligent design and construction of underground storage, especially salt caverns; and (vi) intelligent monitoring, optimization, and risk management. Despite these advances, challenges persist in data scarcity, physical consistency, and generalization. Future efforts should focus on hybrid physics-informed AI, digital twin-enabled operation, and multi-gas comparative frameworks to achieve safe, efficient, and intelligent underground storage systems aligned with global carbon neutrality.

1. Introduction

The global energy system is undergoing profound transformation driven by decarbonization, electrification, and large-scale renewable integration [1,2]. Underground gas storage (UGS) technologies—including hydrogen storage in porous media [3] and salt caverns [4], seasonal natural gas storage in depleted hydrocarbon reservoirs and salt caverns [5], compressed air energy storage (CAES) in salt caverns [6,7], and lined rock caverns [8], as well as repurposed abandoned mines [9] (Figure 1)—are increasingly recognized as indispensable solutions for grid balancing and long-term security of supply [10]. Nevertheless, the design and operation of UGS systems are complicated by geological heterogeneity, multiphase flow dynamics, thermo-mechanical interactions, and long-term integrity risks.
Experts and scholars both in China and abroad have conducted extensive research on key scientific and technological issues of UGS engineering, such as reservoir stability, sealing performance, and thermal processes. The mechanical design of salt caverns for gas storage requires careful consideration of the time-dependent creep behavior of rock salt and operational pressure limits. Staudtmeister and Rokahr [13] established the “New Hannover Dimensioning Concept”, integrating laboratory data, finite element modeling, and engineering practice from more than 50 caverns to provide reliable design criteria for long-term stability. Yang et al. [1,14,15] systematically reviewed the fundamental concepts and core challenges of deep underground energy storage technologies, focusing on the key scientific issues faced by UGS and other subsurface storage approaches. They conducted an in-depth analysis of the current development status, China’s research progress in this field, and the future development trends and critical directions. Their work provides a systematic technological roadmap and strategic guidance for the large-scale deployment and industrialization of deep underground energy storage in China. Recent work by Körner et al. [16] employed a novel experimental setup together with numerical modeling to investigate thermally induced fractures in rock salt, demonstrating that gas pressure and rapid cooling jointly contribute to fracture initiation and propagation around storage caverns. Mohanty and Vandergrift [17] proposed a novel boundary-element modeling approach incorporating pillar weakening to evaluate the long-term stability of an old underground propane storage cavern. Their results demonstrated that the cavern could remain serviceable for at least 30 more years, provided pressure conditions are carefully managed. Large-scale underground hydrogen storage (UHS) in porous media has been recognized as a promising solution for seasonal energy balancing, yet its maturity remains low. Heinemann et al. [18] outlined the key scientific challenges—including multiphase flow behavior, geochemical reactions, microbial activity, and geomechanical responses—that must be addressed to ensure safe and efficient implementation. Recent investigations highlight the crucial role of cushion gas in underground hydrogen storage (UHS). Kanaani et al. [19] demonstrated through reservoir-scale simulations that the choice and composition of cushion gas strongly affect hydrogen recovery, with methane showing the most favorable performance in depleted oil reservoirs. Khaledi et al. [20] employed an elasto-viscoplastic creep model to predict the stress–strain relationship of the surrounding rock in salt caverns during construction and cyclic operation for CAES, and carried out numerical simulations of excavation processes and cyclic loading. Berest et al. [5,21] discussed several aspects of the thermodynamic behavior of salt caverns used for CAES and developed a simplified model describing the pressure and temperature in the wellbore during gas withdrawal. Shallow-buried lined rock caverns have been proven feasible for CAES, provided that gas leakage is effectively controlled and structural integrity is maintained. For instance, Kim et al. [22] used coupled simulations to evaluate gas tightness and energy balance under different design scenarios, and further explored the impact of the excavation damaged zone (EDZ) on the geomechanical performance and stability of such caverns [23]. Menéndez et al. [24] investigated the application of abandoned mines for CAES, presenting a case study that integrates CAES, underground pumped hydro storage, and geothermal applications within a single mine.
With the advancement of energy transition and intelligent development, artificial intelligence (AI) technologies are being deeply integrated into the entire process of UGS construction and operation, bringing revolutionary opportunities to enhance exploration accuracy, construction efficiency, and intelligent decision-making in underground storage facilities. AI, leveraging machine learning (ML), deep learning (DL), and data-driven analytics, has demonstrated remarkable potential in solving such complex nonlinear problems. Recent progress has highlighted the role of AI in optimizing complex UGS processes. Aghdam et al. [25] proposed a hybrid workflow that integrates machine-learning-based surrogate models with multi-objective optimization to design cushion gas strategies in depleted oil reservoirs, simultaneously enhancing gas recovery and CO2 sequestration The combination of AI and geomechanics is gaining attention. Lu et al. [26] demonstrated this by employing a transformer-based model for long-term prediction of thermo-mechanical responses in cryogenic storage, highlighting the importance of integrating field data, numerical modeling, and ML. Given the vast complexity and uncertainty in distinguishing induced from natural earthquakes, Foulger et al. [27] emphasized the urgent need for more systematic hazard assessment in underground storage projects. This provides a strong rationale for applying AI techniques—such as ML for seismic pattern recognition and risk prediction—to enhance early warning and decision-making. Chellal et al. [3] developed a deep-learning-based surrogate modeling framework for fault reactivation risk assessment in underground hydrogen storage, demonstrating that AI can significantly improve the reliability of geomechanical risk predictions compared to conventional analytical models. In recent years, ML approaches have been applied to forecast the wetting characteristics and interfacial interactions between rocks and fluids [28,29,30,31]. To address the challenges of characterizing wettability in hydrogen storage systems, Turkson et al. [32] introduced HyWEC, an interpretable machine-learning framework that enables accurate and transparent estimation of hydrogen wettability, reducing reliance on resource-intensive laboratory experiments. Accurate estimation of interfacial tension (IFT) is vital for understanding hydrogen migration and storage efficiency. Tackie-Otoo et al. [33] demonstrated that integrating Bayesian optimization with advanced ML algorithms can significantly improve IFT prediction accuracy in quaternary hydrogen systems, thereby reducing reliance on costly laboratory experiments. Building on emerging AI approaches for pore-scale property prediction, recent works have shown that machine-learning models can deliver accurate and efficient estimates of hydrogen–brine IFT—a key control on capillary entry and storage performance [34]. Using interpretable ANFIS-GA schemes and high-accuracy RBFNNs optimized by swarm-intelligence algorithms, these studies markedly reduce experimental burden and broaden operational envelopes for underground hydrogen storage design [35].
Despite the growing body of research on AI-assisted underground gas storage (UGS), several critical research gaps remain. First, most existing studies are confined to a single storage type—such as salt caverns, depleted reservoirs, or aquifers—without conducting systematic cross-type comparisons. This limits the generalization of AI models and their transferability across different geological and operational contexts. Second, research on collaborative optimization and mutual adaptation among multiple gas types (e.g., hydrogen, natural gas, and compressed air) is still scarce, even though these systems share similar thermodynamic and operational principles. Third, current studies seldom establish unified AI frameworks that integrate multi-source geological, thermodynamic, and operational data to enable cross-gas learning and decision-making. Addressing these deficiencies is essential to transition from fragmented, case-specific applications toward a comprehensive, interoperable, and multi-gas intelligent paradigm for UGS engineering. Therefore, this review aims to bridge these gaps by performing cross-type comparative analyses and exploring the prospects of AI-driven collaborative optimization for multi-gas underground storage systems.
This study centers on the applications and challenges of AI in UGS, with the aim of advancing intelligent, resilient, and efficient design, construction, and operation of storage facilities in the context of the global energy transition. This paper aims to provide a comprehensive bibliometric review of AI applications in UGS, integrating hydrogen, natural gas, and CAES within a unified framework. Using knowledge-graph methods, we identify research hotspots, methodological advances, and future directions. In Section 2, by establishing a bibliometric analysis framework based on the Web of Science Core Collection (WoSCC) database and employing VOSviewer (version 1.6.20), as well as CiteSpace (version 6.4.R2) for knowledge-graph visualization, this study systematically identifies the research hotspots and development trends of AI applications in UGS. Building upon this, in Section 3, it further synthesizes the technological pathways and representative case studies of AI across multiple application scenarios—including site characterization and modeling, intelligent construction, stability prediction, operational optimization of injection–production cycles, and safety monitoring. Considering the current challenges, the study proposes future directions such as integrating physics-informed AI, constructing comprehensive engineering data platforms, and developing interoperable intelligent systems, with the aim of providing theoretical foundations and technical references for the intelligent transformation and sustainable development of UGS engineering.

2. Methods

2.1. Data Collection

To explore the latest developments of AI in UGS systems, this study used the WoSCC as the data source. Relevant keywords, abstracts, and other bibliographic information were extracted through topic searches based on logical operators “AND” and “OR”, with the search codes and steps illustrated in Figure 2. The dataset was retrieved from the WoSCC using the query as fellow:
TS = (“xxx*” AND/OR “xxx*”)
Here, ‘TS’ denotes the topic field, quotation marks indicate exact phrase searching, ‘xxx’ represents the search term, and the asterisk ‘*’ is used as a wildcard for fuzzy matching.
The retrieval process was carried out in three stages (Figure 2): in Step 1, papers published between 1 January 1900, and 31 August 2025, on the topic of UGS were collected. In Step 2, the initial results were refined by combining them with AI-related topic terms to optimize the search. In Step 3, visualization analyses were conducted using VOSviewer and CiteSpace on the selected publications from the WoSCC to infer research hotspots and trends of AI applications in UGS systems. Furthermore, a critical review of the literature was carried out to summarize the current state of AI development in this field. It should be noted that the search term “neural network” encompasses artificial neural networks (ANNs), convolutional neural networks, recurrent neural networks, and physics-informed neural networks.

2.2. Knowledge Graph and Bibliometric Analysis

Based on the Step 1 search results shown in Figure 2, the temporal evolution of global publications on UGS research was plotted, as illustrated in Figure 3a. The evolution of global publications on UGS research reveals three distinct phases. In the early stage (before the 1990s), research output remained sparse. A slow growth phase followed between the 2000s and mid-2010s. Since 2015, the field has entered a rapid growth phase, a phase of diversification and intelligent development, marked by a surge of studies on various geological formations—including salt caverns, depleted hydrocarbon reservoirs, abandoned mines, and lined rock caverns—and the increasing integration of AI and ML. These trends underscore the growing strategic importance of UGS in supporting renewable energy integration and global carbon neutrality goals.
Based on the Step 2 search results shown in Figure 2, the temporal evolution of global publications on AI in UGS research was plotted, as illustrated in Figure 3b. The annual publication trend on AI applications in UGS technologies exhibits a steady growth, indicating increasing scholarly attention to this emerging interdisciplinary field. The evolution can be divided into two phases: an initial exploratory stage (before 2017) with limited publications, a growth phase (since 2017) driven by the integration of ML into geomechanics and fluid modeling, and a recent surge characterized by deep learning, proxy modeling, and digital twin applications, particularly in hydrogen storage. This shift reflects both the global momentum of AI after 2017 and the growing demand for intelligent solutions in hydrogen storage and CAES under the carbon neutrality agenda.
Based on the 3771 publications retrieved from Step 1 in Figure 2, VOSviewer was employed to conduct a keyword co-occurrence visualization analysis. This approach allows for the identification of research hotspots, thematic structures, and emerging trends in the application of AI within UGS studies. In VOSviewer, keywords with at least 15 occurrences were selected, resulting in 283 terms. Higher thresholds ensured that only representative keywords with strong connectivity were retained for global UGS trends, whereas a lower threshold captured emerging topics in the smaller AI-UGS dataset (176 records). Among these, the top 283 keywords ranked by total link strength were retained for co-occurrence network analysis, as shown in Figure 4a. Keyword co-occurrence analysis (Figure 4a) reveals that research on UGS can be divided into five thematic clusters, each representing a distinct research theme in UGS. The clusters are color-coded into red, green, blue, yellow, and purple, reflecting different topical focuses such as thermodynamic and multiphase flow modeling, geomechanics and stability evaluation, hydrogen and CO2 storage processes, operational optimization and monitoring, and renewable energy integration. Within this network, the “machine learning” node appears in the red cluster. Although its relative size remains modest (94 occurrences, 236 total link strength), it establishes strong connections with frontier research topics such as underground hydrogen storage (233 occurrences, 2274 link strength), CO2 storage (180 occurrences, 566 link strength), pressure (222 occurrences, 1189 link strength), and interfacial tension (62 occurrences, 189 link strength). This indicates that AI methods are increasingly being coupled with core challenges in UGS, particularly predictive modeling of storage performance, interfacial property estimation, and subsurface pressure evolution.
The statistics further reveal a temporal dimension: while most geomechanics-related keywords (e.g., stability, creep, rock salt) gained prominence before 2017, AI-related terms (e.g., ML, algorithm, prediction) became more visible after 2017, consistent with the broader rise of AI in geoenergy. Notably, underground hydrogen storage shows a steep growth trajectory (233 occurrences, 2023.5 average publication year), highlighting its status as the fastest-emerging theme where AI plays a pivotal supporting role.
Overall, the co-occurrence map demonstrates that while AI is not yet the dominant driver in UGS research, it is forming bridging links to multiple frontier areas, suggesting a shift toward AI-assisted site characterization, proxy modeling, and intelligent monitoring. This trajectory positions AI as a transformative enabler for next-generation UGS systems.
Based on the 176 publications retrieved from Step 2 in Figure 2, VOSviewer was employed to conduct a keyword co-occurrence visualization analysis. In VOSviewer, keywords with at least 5 occurrences were selected, meaning that only keywords appearing at least five times in the dataset were included in the analysis. Out of a total of 950 extracted keywords, 44 terms met this threshold, which were further grouped into four clusters (Figure 4b).
The red cluster mainly contains terms such as AI, ANN, algorithm, model, and optimization, reflecting the methodological foundation of AI applications in UGS. The green cluster includes machine learning, renewable energy, porous media, flow, and solubility, which highlight the role of AI in energy systems modeling, porous media characterization, and multi-energy integration. The blue cluster consists of keywords such as hydrogen storage, pressure, temperature, IFT, and wettability, focusing on the prediction of subsurface processes and thermo-hydraulic mechanisms in hydrogen and natural gas storage. The yellow cluster, centered on prediction and storage, represents the direct application of AI methods in performance forecasting, system reliability assessment, and storage optimization.
Overall, ML emerges as the most influential node, with the highest occurrence (52) and link strength (171), serving as a pivotal bridge between computational intelligence and subsurface engineering applications. Meanwhile, keywords such as hydrogen storage and underground hydrogen storage demonstrate that AI is increasingly embedded in the emerging field of hydrogen-based geoenergy systems. This clustering result suggests that AI has become a critical enabler in linking advanced algorithms with complex multi-physical processes, thereby supporting the intelligent and sustainable development of underground storage technologies.
Based on the keywords in 176 publications retrieved from Step 2 in Figure 2, CiteSpace was employed to generate the timeline visualization (Figure 5) and mainly found 9 burst words (Figure 6). The timeline knowledge-graph in Figure 5 was generated using CiteSpace (version 6.4.R2). The analysis was performed on the Step 2 dataset (176 publications) over the timespan 1990–2025 with a slice length of 1 year. The node type was set to “keyword”, and clustering was conducted using the g-index as the selection criterion (k = 25; LRF = 2.5; L/N = 10; LBY = 5; e = 1.0), without additional pruning. The resulting clusters (#0–#9) exhibited a modularity Q of 0.5877 and a weighted mean silhouette S of 0.8315, indicating a well-partitioned and internally coherent community structure. Figure 5 highlights the temporal evolution of research topics in AI applications for UGS. Each cluster (#0–#9) represents a thematic focus, while the horizontal axis indicates the time span of keyword activity. Cluster #0 “artificial intelligence” emerges as the largest and most sustained research front, showing continuous development from 2018 onwards. This cluster includes keywords such as neural networks, optimization, and simulation, reflecting the methodological core of AI in UGS studies. Cluster #1 “underground natural gas storage” and Cluster #3 “gas storage” emphasize engineering practices and site-specific issues, indicating strong ties between AI methods and traditional UGS engineering problems such as injection–production strategies and reservoir performance evaluation. Cluster #2 “artificial neural networks” is closely associated with early stage applications of AI, particularly in predictive modeling and system performance evaluation. Its prominence before 2020 suggests that ANN was the primary entry point for AI into UGS research. Cluster #4 “renewable energy” and Cluster #5 “energy transition” appear more recently (post-2020), reflecting the integration of UGS with broader decarbonization strategies, where AI supports optimization and multi-energy system coordination. Cluster #6 “hydrogen solubility” and Cluster #7 “hydrogen storage” show the emerging role of AI in hydrogen-related subsurface processes, such as predicting gas–brine interactions, wettability, and IFT—topics that gained momentum particularly after 2021. Clusters #8 “solar assisted” and #9 “risk assessment” represent niche but growing themes, addressing hybrid renewable integration and safety management of UGS systems. Overall, the timeline view demonstrates a shift from early applications of ANN and optimization (2018–2020) toward more complex AI-driven studies on hydrogen storage and energy transition after 2020. This indicates that AI research in UGS has moved from methodological exploration to problem-oriented applications aligned with global carbon neutrality and renewable integration goals.

3. Typical Application Scenarios of AI in UGS Engineering

3.1. Intelligent Characterization of Rock Mechanical Parameters for UGS

For intelligent characterization of rock mechanical parameters relevant to UGS, using nanoindentation and SEM-EDS, a multi-parameter dataset was built to train XGBoost predictors of shale hardness and reduced modulus from high-resolution SEM images [36]. Mineral-phase behaviors were separated via XRD composition and GMM clustering, and Mori–Tanaka upscaling yielded homogenized Young’s modulus and Poisson’s ratio at the centimeter scale. The approach provides essential parameters for energy applications and a generalizable reference for property estimation in other materials. Mohanto and Deb [37] developed multivariate regression and ANN models to predict the plastic damage index of rib pillars in underground metal mines, providing a novel AI-based approach for rock mass stability assessment. Shin [38] applied supervised ML to classify lithologies in a vanadiferous titanomagnetite deposit, achieving high-accuracy integration of multi-geophysical data for deep ore body prediction. Recent advances in salt rock mechanics have focused on accurate constitutive modeling and parameter calibration under cyclic loading conditions. Honório et al. [39] developed a comprehensive multi-step calibration strategy combining elastic, viscoelastic, viscoplastic, and dislocation-creep mechanisms, enabling reliable determination of representative parameters for salt rocks using Particle Swarm Optimization (PSO). This approach ensures robustness against sample heterogeneity and progressively improves model accuracy as new experimental data are incorporated. In contrast, He et al. [40] experimentally investigated the fatigue–creep coupling behavior of Jintan salt rock under complex cyclic loading and established a fatigue–creep composite damage model based on the Burgers framework, as shown in Figure 7. Using ML (XGBoost + SHAP), they quantified the nonlinear influence of stress magnitude and holding time on fatigue life. Their regression model achieved coefficients of determination of R2 = 0.89 and 0.81 for predicting fatigue life (N) and peak strain (ε1max), respectively, demonstrating the capability of AI algorithms to accurately characterize complex nonlinear relationships among multiple mechanical factors. Together, these studies complement each other: the former advances parameter identification and calibration methodology, while the latter deepens understanding of damage evolution mechanisms, offering a foundation for integrated, AI-assisted constitutive modeling of salt caverns under high-frequency injection–production cycles.
Building on previous developments in pore-scale imaging and digital rock reconstruction, several studies have explored advanced AI-based methods for microscopic reservoir characterization. Guan et al. [41] comprehensively reviewed microscopic characterization methods and fractal analysis of unconventional reservoir pore systems, highlighting the correlation between pore–throat complexity and reservoir properties, and suggesting future integration with AI and 4D imaging techniques. Zhu et al. [42] proposed a diffusion model-based ML method for generating three-dimensional multiphase pore-scale images, enabling realistic reconstruction of rock pore structures and fluid distributions for applications such as multiphase flow and underground hydrogen storage. Deep learning models such as autoencoder and U-Net can accurately predict fine particle transport and retention in porous media, achieving SSIM values up to 0.95 and R2 values of 0.97 for porosity and 0.88 for permeability [43]. Compared with the work of Maleki et al. [44], which applied CatBoost and Extra Trees to model CO2 and CH4 adsorption in tight rocks, the study of Kalam et al. [45] extended the application of ML to hydrogen adsorption on shale kerogen, confirming that gradient boosting regression can reach comparable accuracy while significantly reducing computational time.

3.2. Multiphysics Coupling and Surrogate Modeling

Recent advances in AI-driven modeling have significantly improved the efficiency and accuracy of subsurface energy storage simulations. Neural operator-based frameworks (e.g., FFINO (factorized Fourier improved neural operator), UHSNet, and Gazprom’s tensor-decomposed FNO (Fourier neural operator)) have demonstrated orders-of-magnitude acceleration in solving multiphase flow equations, enabling real-time prediction and optimization of underground hydrogen or gas storages. The FFINO architecture was developed as a factorized Fourier neural operator that integrates relative permeability uncertainty into operator learning, achieving up to 7850× faster inference and nearly 10% higher accuracy than conventional models in simulating multiphase hydrogen–brine flow for underground hydrogen storage systems [46]. The UHSNet model was introduced as a smart deep learning proxy framework employing dilated convolutional layers and a hybrid Huber–MAE loss function, achieving a 15% reduction in prediction error and operating 104× faster than CFD simulations for predicting hydrogen plume evolution in saline aquifers [47]. A tensor-decomposed FNO was proposed for hydrodynamic modeling of UGSs, achieving a 50× reduction in trainable parameters and enhanced generalization to complex geometries while maintaining accuracy comparable to numerical solvers [48].
In parallel, hybrid optimization algorithms such as HGA–GRG (hybrid genetic algorithm–generalized reduced gradient) enhance surface system efficiency, while physics-informed ML frameworks extend AI applications to coupled thermal–fluid processes in large underground systems. A HGA–GRG framework was established to optimize injection and production pipeline parameters under harmonized operating conditions, resulting in an investment cost reduction of over CNY 3.45 million and a significant acceleration in convergence compared with traditional genetic algorithm (GA) methods [49]. A physics-informed ML framework combining LSTM–Seq2Seq architecture and attention mechanisms was proposed for real-time prediction of air temperature and humidity in large underground tunnels, achieving temperature errors below 0.4 °C and humidity errors below 2%, thereby enabling active thermal management of subsurface spaces [50]. Wang et al. developed a gradient-boosted spatiotemporal neural network (GSTNN) that integrates governing gas–water seepage and convection–diffusion equations into the neural network structure by incorporating partial differential operators as regularization terms in the loss function, thereby achieving physics-consistent simulation of underground hydrogen storage processes [51], as shown in Figure 8. Collectively, these studies mark a paradigm shift toward intelligent, physics-consistent, and computationally efficient modeling of underground energy systems.

3.3. Gas–Rock–Fluid Interaction, Wettability and Interfacial Behavior Prediction

AI has also found typical applications in addressing geochemical interface issues in gas storage. By integrating ML with molecular simulations and experimental data, predictive models for IFT, wettability, and solubility have been developed to tackle challenges related to phase behavior, leakage risk, and storage efficiency during hydrogen or natural gas storage in salt rocks and saline aquifers. In recent years, the incorporation of physics-informed neural networks and Bayesian optimization has further enhanced model generalization and interpretability. Multiple studies have employed ML and Bayesian optimization methods to predict the IFT of hydrogen–brine systems, thereby improving the prediction accuracy of capillary entry pressure and gas storage performance in underground hydrogen storage [34,52,53,54,55]. The interpretable machine-learning tool HyWEC, combined with Bayesian-optimized models, has been applied to predict the IFT and wettability of hydrogen–brine systems, markedly enhancing the precision of capillary entry pressure estimation and storage performance evaluation in underground hydrogen storage [32,33]. A series of recent studies have systematically applied ML to model interfacial and transport properties in hydrogen–brine systems. Support vector regression has been used to predict gas dispersion coefficients in porous media with near-perfect accuracy (R2 > 0.99) [56], while ensemble learning models such as XGBoost and LightGBM enable precise estimation of hydrogen wettability and column height in various rock–brine systems [57]. More advanced frameworks, including the RBFNN (radial basis function neural network)–AGTO (artificial gorilla troops optimizer) hybrid [35] and the interpretable GMDH (group method of data handling)/GEP (gene expression programming) white-box models [58] have achieved unprecedented precision in IFT prediction and provided explicit correlations for underground hydrogen storage design, as shown in Figure 9. ML models, including multilayer perceptron, genetic programming, and the group method of data handling, were applied to predict hydrogen solubility in brine under high-pressure and high-temperature conditions, with the multilayer perceptron optimized by the Levenberg–Marquardt algorithm achieving superior accuracy (R2 = 0.9991, AARE = 0.94%), highlighting the potential of ML approaches for simulating dissolution processes in underground hydrogen storage [59].

3.4. Injection-Production Process Optimization

Optimization of injection–production operations, pressure differential control, and energy utilization efficiency in UGS well clusters has attracted increasing attention from researchers. AI techniques—including GAs, reinforcement learning, deep reinforcement learning, and proxy models—have been employed to develop integrated surface–subsurface optimization frameworks aimed at achieving objectives such as injection–production flow optimization, pipeline network scheduling, and minimization of energy losses. Some studies have further incorporated Digital Twin technology to enable real-time pressure prediction and autonomous regulation. Huohai et al. [60] proposed a hybrid agent-based optimization framework for UGS injection–production operations, in which nine ML models were compared and the XGBoost-based surrogate model was identified as the most accurate predictor. By integrating this model with the NSGA-III multi-objective optimization algorithm, the framework achieved a 64% reduction in inter-block pressure difference and enhanced storage efficiency, demonstrating the feasibility of AI-driven operational optimization in UNGS systems. Li et al. [61] developed a two-layer co-optimization framework for underground natural gas storage that integrates aboveground compressor operation with underground injection scheduling. Using a ML-based proxy model, the method reduced energy consumption by 10.79% and reservoir pressure deviation by 85.62%, demonstrating its effectiveness for safe and efficient injection management. Kuk et al. [62] proposed an auto-adaptive decision tree method that integrates ML and optimal control theory for intelligent well operation in high-nitrogen UGS. The adaptive model reduced nitrogen content by 2.4% without increasing costs, demonstrating the potential of AI-based well control to enhance energy efficiency and gas quality. Indro et al. [63] used a deep neural network-based reduced-order model trained on 1200 simulations to optimize multi-well strategies for underground hydrogen storage in depleted gas reservoirs, as shown in Figure 10. The study showed that multi-well configurations and optimal well spacing markedly improve hydrogen purity, capacity, and efficiency, highlighting the value of ML for UHS design and operation.
Güyagüler and Gümrah [64] applied GA to optimize gas production rates in UGS systems, marking one of the earliest uses of AI in subsurface storage optimization. The authors developed an interactive program, GASOPT, which couples a three-dimensional gas reservoir simulator with GA to maximize cumulative gas production under operational constraints such as minimum flowing well pressure. Compared with traditional linear programming (LP) approaches, the GA-based method avoided excessive linearization assumptions and provided more accurate and realistic optimization results. This pioneering study laid the groundwork for subsequent AI-driven optimization of underground storage systems. A GA-based optimization approach was developed to replace traditional LP methods for underground storage field optimization [65]. Unlike LP, which relies on multiple linearization assumptions that reduce accuracy, the GA directly couples with a 3D finite-difference reservoir model, evaluating well rate configurations through evolutionary selection. This method achieved more realistic and accurate optimization results, outperforming LP due to its ability to handle nonlinear field behavior.

3.5. Monitoring, Risk Assessment and Integrity Evaluation

AI and sensing technologies have significantly enhanced the monitoring and management of UGS systems. Wamriew et al. [66] integrated distributed acoustic sensing (DAS) with deep learning to achieve real-time detection and localization of microseismic events during reservoir operations. Using convolutional neural networks (CNNs), their framework successfully identified subtle seismic signals from massive DAS datasets, providing high spatial–temporal resolution for fracture characterization in CO2 sequestration and hydrogen storage applications. Building on the potential of DAS for intelligent wellbore surveillance, Su et al. [67] introduced a CWT–CNN–BiLSTM model capable of capturing non-stationary acoustic features associated with cement-sheath degradation. The system achieved 99.31% classification accuracy in detecting top-of-cement and microcrack leakage failures, establishing a real-time and full-cycle evaluation framework for wellbore integrity. At the reservoir scale, Huang et al. [68] proposed an InSAR-based hybrid prediction model combining the Gray Wolf Optimization, Variational Mode Decomposition, and Gated Recurrent Unit (GWO–VMD–GRU) algorithms to forecast long-term surface deformation at the Hutubi UGS in Xinjiang, China. The model achieved an R2 greater than 0.98 and revealed clear correlations between gas injection–production cycles and ground displacement, providing a robust AI-driven approach for operational safety prediction and management. Zhang et al. [69] developed an improved Mask R-CNN model with attention mechanisms and a hybrid loss function for high-precision crack detection in shallow-buried CAES cavern linings, achieving a mean average precision of 89.3% and providing a new intelligent approach for cavern safety assessment. Meanwhile, He et al. [70] constructed a large-scale permeability dataset of rock–concrete interfaces under cyclic loading and proposed an LSTM-SSA intelligent prediction framework that accurately captures seepage evolution and identifies key influencing factors, offering guidance for the long-term sealing design of artificial CAES caverns.
AI has enabled transformative approaches for monitoring, control, and safety assurance in underground energy storage systems. Gutiérrez-Oribio et al. [71] proposed a reinforcement-learning-based robust control framework to optimize subsurface energy extraction while minimizing induced seismicity. By coupling Biot’s poroelastic model with a deep deterministic policy gradient (DDPG) controller, their system dynamically adjusts injection rates and control gains to achieve a balance between sustainable production and seismic hazard mitigation, establishing a novel “control–learning” hybrid paradigm for safe operation of geothermal and hydrogen storage reservoirs. In parallel, Sugan et al. [72] developed the LOC-FLOW workflow integrating the deep-learning model PhaseNet into microseismic monitoring of the Collalto UGS in Italy. The system automatically detects and locates more than 90% of seismic events compared with manual catalogs, reducing human intervention and enabling near-real-time monitoring of induced seismicity. Complementing these efforts, Gao et al. [73] introduced HydrogenNet, a supervised machine-learning model combining 2D convolutional neural networks (CNNs) and dense neural networks (DNNs) for detecting and characterizing hydrogen leakage from underground hydrogen storage using sparse time-lapse seismic data. The model achieves high accuracy in estimating leakage location, mass, and volume even under noisy conditions, providing a cost-effective solution for subsurface integrity assessment.
Recent advances have leveraged ML to enhance the prediction of UGS deliverability. Vo Thanh et al. [74] compared GPR, LSSVM, and Extra Tree models across 387 storage sites, revealing that GPR provided the highest accuracy (R2 ≈ 1.0) for various geological formations. Building on this, Wei et al. [75] developed a stacking ensemble framework combining XGBoost, LightGBM, and CatBoost, achieving R2 > 0.9999 and improved interpretability via SHAP analysis. Together, these studies highlight the growing role of AI-driven models in reliable, data-informed evaluation of storage capacity and operational performance. Derakhshani et al. [76] developed an AI-based framework integrating ML, MCDA, and GIS to assess hydrogen storage potential in Poland’s bedded salt formations. Among eight tested algorithms (KNN, SVM, LightGBM, XGBoost, MLP, CatBoost, GBR, and MLR), CatBoost achieved the highest accuracy (R2 = 0.888), highlighting storage capacity and energy demand as key factors. The approach offers a fast and transferable method for identifying optimal UHS sites using ensemble learning. The growing potential of AI-driven models—from feed-forward ANN to transformer architectures—has been demonstrated in efficiently capturing complex long-term mechanical and thermo-mechanical behaviors in underground energy storage systems. Zhao et al. [77] developed a deep learning approach using an ANN to predict the long-term stability of energy-storage salt caverns, achieving coefficients of determination above 0.97 for wall displacement and volume reduction, thereby offering a rapid alternative to time-consuming numerical simulations. In comparison, Lu and Wu [26] applied a transformer-based model trained on coupled stress–temperature-gradient experiments to forecast 210-day thermo-mechanical responses around cryogenic storage caverns, outperforming conventional numerical methods in replicating field temperature and deformation trends, as shown in Figure 11.
A digital twin-based framework [78] has been developed to evaluate and predict well integrity in UGS systems by coupling material and energy balance models with GA optimization, enabling accurate identification and quantification of leaks through real-time temperature and pressure trend analysis, thereby enhancing operational safety and reducing intervention costs.

3.6. Intelligent Design and Construction of Underground Storage Cavern

To achieve intelligent design and efficient construction of UGS facilities, particularly salt caverns, recent studies have introduced ML approaches such as ANNs, significantly improving the efficiency and accuracy of cavern geometry prediction, capacity estimation, and parameter optimization [79,80,81,82,83]. Among these, a geometry prediction system based on the Gated Recurrent Unit (GRU) model can directly predict cavern radii at different heights from multi-stage leaching parameters, enabling rapid geometric inversion and target-shape construction [79]. The model achieved a mean absolute error of only 1.6 m in the test dataset. Meanwhile, a back propagation neural network has been applied to capacity and maximum radius prediction, enabling the rapid screening of over one million sets of design parameters within less than one second—approximately 6 × 107 times faster than traditional numerical simulations [80], as shown in Figure 12. Furthermore, to address the challenges of multiphase flow simulation in the salt dissolution process, researchers have developed an ANN-based Volume-of-Fluid interface normal prediction model, which substantially enhances the accuracy of three-dimensional interface reconstruction and provides crucial boundary support for intelligent cavern modeling [81]. These advances indicate that AI-based methodologies are gradually forming a closed-loop “data–model–optimization” framework, driving salt cavern engineering toward higher efficiency, precision, and intelligence.
To provide a concise synthesis of the six AI application scenarios discussed above (Section 3.1,Section 3.2,Section 3.3,Section 3.4,Section 3.5 and Section 3.6), Table 1 presents a synthesized framework highlighting the AI task categories, commonly used algorithms, data sources, and major limitations across different UGS application scenarios.

4. Conclusions

This review provides a systematic and comprehensive overview of the integration of AI into UGS engineering by combining bibliometric and knowledge-graph analyses with an in-depth technical review, bridging disciplines such as geomechanics, thermodynamics, fluid dynamics, and data science. The integration of AI has reshaped the entire UGS lifecycle—from geological characterization and construction design to monitoring, optimization, and long-term integrity evaluation. Technically, hybrid physics-informed AI refers to the coupling of data-driven algorithms (e.g., deep neural networks) with physical and thermodynamic constraints derived from governing equations, allowing AI models to capture both data patterns and mechanistic laws. Likewise, a multi-gas comparative framework envisions a unified architecture capable of transferring learning and parameterization across hydrogen, natural gas, and compressed-air storage systems, thereby enabling collaborative optimization under shared geomechanical and thermodynamic principles. The following conclusions were drawn:
(1)
Since 2017, AI research in UGS has rapidly transitioned from traditional ANNs toward advanced deep learning and physics-informed frameworks, with underground hydrogen storage emerging as the dominant frontier.
(2)
ML models such as CNN–LSTM, transformer networks, and FNO-based operators have enabled orders-of-magnitude acceleration in multiphysics simulations, enhanced predictive accuracy in interfacial behavior, and improved real-time monitoring capabilities.
(3)
AI-driven optimization frameworks—integrating reinforcement learning, GAs, and digital twin systems—have achieved measurable gains in operational efficiency, energy utilization, and safety reliability across salt caverns, depleted reservoirs, and lined rock caverns.
(4)
Despite promising progress, AI applications still face challenges related to data scarcity, geological heterogeneity, and physical interpretability. Addressing these requires the fusion of multi-source data, domain knowledge, and physics-guided constraints.
Looking ahead, the convergence of AI, digital twin, and advanced sensing technologies will accelerate the realization of self-adaptive and resilient underground storage systems. AI is no longer an auxiliary analytical tool but a transformative driver for next-generation UGS engineering—enabling data-driven, physically consistent, and sustainable subsurface energy management. In particular, physics-informed AI and cross-gas comparative frameworks will be key to advancing intelligent underground hydrogen and compressed air storage, contributing to the global energy transition and carbon neutrality.
Furthermore, several methodological limitations are acknowledged. The bibliometric analysis is inherently affected by database coverage bias (restricted to the WoSCC), language bias (English-dominated publications), and keyword selection bias (threshold sensitivity in VOSviewer and CiteSpace). Recognizing these constraints provides context for interpreting the quantitative trends and highlights opportunities for future meta-analyses.
Overall, the findings not only summarize existing applications but also provide a synthesized outlook for developing interoperable, physics-informed, and multi-gas intelligent UGS systems.

Author Contributions

Methodology, writing—original draft preparation, software, J.C. and G.W.; validation, funding acquisition, X.B.; formal analysis, project administration, C.D.; investigation, data curation, J.L. (Jun Lu) and L.X.; methodology, writing—review and editing, software, X.G.; supervision, project administration, G.Z.; supervision, visualization, J.L. (Jinlong Li). All authors have read and agreed to the published version of the manuscript.

Funding

The National Pipeline Network Group’s scientific research and technology development project, “Research on Key Technologies for Intelligent Leaching Design of Salt Cavern Gas Storage Facilities”, No.CNJS-SZZN202401.

Data Availability Statement

The original contributions presented in this study are included in the article. Further inquiries can be directed to the corresponding author.

Acknowledgments

We would like to express our sincere gratitude to PipeChina Corporation for its support of the intelligent construction project for underground gas storage facilities.

Conflicts of Interest

Authors Jiasong Chen, Guijiu Wang, Xuefeng Bai, Chong Duan, Jun Lu and Luokun Xiao were employed by the company Pipechina Energy Storage Technology Co., Ltd. The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Abbreviations

The following abbreviations are used in this manuscript:
UGSUnderground gas storage
AIArtificial intelligence
CAESCompressed air energy storage
MLMachine learning
IFTInterfacial tension
GAGenetic algorithm
ANNArtificial neural network
WoSCCWeb of Science Core Collection

References

  1. Yang, C.H.; Wang, T.T.; Chen, H.S. Theoretical and technological challenges of deep underground energy storage in China. Engineering 2023, 25, 168–181. [Google Scholar] [CrossRef]
  2. Ali, M.; Isah, A.; Yekeen, N.; Hassanpouryouzband, A.; Sarmadivaleh, M.; Okoroafor, E.R.; Al Kobaisi, M.; Mahmoud, M.; Vahrenkamp, V.; Hoteit, H. Recent progress in underground hydrogen storage. Energy Environ. Sci. 2025, 18, 5740–5810. [Google Scholar] [CrossRef]
  3. Chellal, H.; Malki, M.; Indro, A.; Mao, S.; Min, W.; Egenhoff, S.; Mehana, M. A quantitative risk assessment framework for fault reactivation in underground hydrogen storage: Coupled simulation and deep learning approach. J. Rock Mech. Geotech. Eng. 2025, in press. [Google Scholar] [CrossRef]
  4. Wei, X.; Shi, X.; Ma, H.; Ban, S. A comprehensive feasibility evaluation of hydrogen-carbon-salt cavern-renewable energy (HCSR) system. Energy 2025, 331, 137040. [Google Scholar] [CrossRef]
  5. Berest, P.; Louvet, F. Aspects of the thermodynamic behavior of salt caverns used for gas storage. Oil Gas Sci. Technol. 2020, 75, 10. [Google Scholar] [CrossRef]
  6. Huang, J.; Ge, X.; Ma, H.; Zhao, T.; Li, Y.; Shi, X. A study on thermodynamic coupling in dynamic injection and production processes of compressed air energy storage. Energy 2025, 319, 135093. [Google Scholar] [CrossRef]
  7. Wan, M.; Ji, W.; Wan, J.; He, Y.; Li, J.; Liu, W.; Jurado, M.J. Compressed air energy storage in salt caverns in China: Development and outlook. Adv. Geo-Energy Res. 2023, 9, 54–67. [Google Scholar] [CrossRef]
  8. Xia, C.C.; Xu, Y.J.; Zhou, S.W.; Qin, S.K.; He, X.H. Fracture initiation and propagation in the lined underground caverns for compressed air energy storage: Coupled thermo-mechanical phase-field modeling. Comput. Geotech. 2023, 157, 21. [Google Scholar] [CrossRef]
  9. Chen, X.; Wang, J.G. Stability analysis for compressed air energy storage cavern with initial excavation damage zone in an abandoned mining tunnel. J. Energy Storage 2022, 45, 103725. [Google Scholar] [CrossRef]
  10. Yang, C.; Wang, T. Underground Rock Salt Used for Energy Storage: Theory and Engineering Practice; Springer Nature: Berlin/Heidelberg, Germany, 2024. [Google Scholar]
  11. Huang, Y.; Shi, X.; Zhu, S.; Wei, X.; Bai, W.; Li, P.; Xu, M.; Li, Y.; Yang, C. Comparative techno-economic analysis of large-scale underground hydrogen storage. Appl. Energy 2025, 400, 126564. [Google Scholar] [CrossRef]
  12. Tackie-Otoo, B.N.; Haq, M.B. A comprehensive review on geo-storage of H2 in salt caverns: Prospect and research advances. Fuel 2024, 356, 129609. [Google Scholar] [CrossRef]
  13. Staudtmeister, K.; Rokahr, R.B. Rock mechanical design of storage caverns for natural gas in rock salt mass. Int. J. Rock Mech. Min. Sci. 1997, 34, 300.e1–300.e13. [Google Scholar] [CrossRef]
  14. Yang, C.; Wang, T. Deep Underground Energy Storage: Aiming for Carbon Neutrality and Its Challenges. Engineering 2023, 29, 11–14. [Google Scholar] [CrossRef]
  15. Yang, C.-H.; Wang, G.-B.; Shi, X.-L.; Zhu, S.-J.; Zheng, Z.-Y.; Liu, W.; Fan, J.-Y. Demands and challenges of large-scale salt cavern hydrogen storage in China. Rock Soil Mech. 2024, 45, 1–19. [Google Scholar]
  16. Körner, F.; Baumgärtel, L.; Leuger, B. Numerical and laboratory investigations of thermally induced fractures in rock salt. IOP Conf. Ser. Earth Environ. Sci. 2023, 1124, 012080. [Google Scholar] [CrossRef]
  17. Mohanty, S.; Vandergrift, T. Long term stability evaluation of an old underground gas storage cavern using unique numerical methods. Tunn. Undergr. Space Technol. 2012, 30, 145–154. [Google Scholar] [CrossRef]
  18. Heinemann, N.; Alcalde, J.; Miocic, J.M.; Hangx, S.J.T.; Kallmeyer, J.; Ostertag-Henning, C.; Hassanpouryouzband, A.; Thaysen, E.M.; Strobel, G.J.; Schmidt-Hattenberger, C.; et al. Enabling large-scale hydrogen storage in porous media—The scientific challenges. Energy Environ. Sci. 2021, 14, 853–864. [Google Scholar] [CrossRef]
  19. Kanaani, M.; Sedaee, B.; Asadian-Pakfar, M. Role of Cushion Gas on Underground Hydrogen Storage in Depleted Oil Reservoirs. J. Energy Storage 2022, 45, 103783. [Google Scholar] [CrossRef]
  20. Khaledi, K.; Mahmoudi, E.; Datcheva, M.; Schanz, T. Stability and serviceability of underground energy storage caverns in rock salt subjected to mechanical cyclic loading. Int. J. Rock Mech. Min. Sci. 2016, 86, 115–131. [Google Scholar] [CrossRef]
  21. Berest, P. Heat transfer in salt caverns. Int. J. Rock Mech. Min. Sci. 2019, 120, 82–95. [Google Scholar] [CrossRef]
  22. Kim, H.M.; Rutqvist, J.; Ryu, D.W.; Choi, B.H.; Sunwoo, C.; Song, W.K. Exploring the concept of compressed air energy storage (CAES) in lined rock caverns at shallow depth: A modeling study of air tightness and energy balance. Appl. Energy 2012, 92, 653–667. [Google Scholar] [CrossRef]
  23. Kim, H.M.; Rutqvist, J.; Jeong, J.H.; Choi, B.H.; Ryu, D.W.; Song, W.K. Characterizing Excavation Damaged Zone and Stability of Pressurized Lined Rock Caverns for Underground Compressed Air Energy Storage. Rock Mech. Rock Eng. 2013, 46, 1113–1124. [Google Scholar] [CrossRef]
  24. Menéndez, J.; Ordóñez, A.; Álvarez, R.; Loredo, J. Energy from closed mines: Underground energy storage and geothermal applications. Renew. Sustain. Energy Rev. 2019, 108, 498–512. [Google Scholar] [CrossRef]
  25. Aghdam, M.A.; Fazaeli, M.; Kanaani, M.; Sedaee, B. Multi-Objective Optimization of Cushion Gas Design in Underground Gas Storage Using Machine Learning. Results Eng. 2025, 27, 106659. [Google Scholar] [CrossRef]
  26. Lu, D.; Wu, W. Long-term thermo-mechanical response of surrounding rock in underground cryogenic storage. Eng. Geol. 2025, 354, 108178. [Google Scholar] [CrossRef]
  27. Foulger, G.R.; Wilson, M.P.; Gluyas, J.G.; Julian, B.R.; Davies, R.J. Global review of human-induced earthquakes. Earth Sci. Rev. 2018, 178, 438–514. [Google Scholar] [CrossRef]
  28. Pan, B.; Song, T.; Yin, X.; Jiang, Y.; Yue, M.; Hoteit, H.; Mahani, H.; Iglauer, S. A Machine Learning Based Framework for Brine-Gas Interfacial Tension Prediction: Implications for H2, CH4 and CO2 Geo-Storage. In Proceedings of the GOTECH, Dubai, United Arab Emirates, 7–9 May 2024. [Google Scholar]
  29. Kim, S.; Kim, T.-W.; Jo, S. Artificial intelligence in geoenergy: Bridging petroleum engineering and future-oriented applications. J. Pet. Explor. Prod. Technol. 2025, 15, 35. [Google Scholar] [CrossRef]
  30. Ali, M.; Tariq, Z.; Mubashir, M.; Kamal, M.S.; Yan, B.; Hoteit, H. In Prediction of Pure Mineral-H2-Brine Wettability Using Data-Driven Machine Learning Modeling: Implications for H2 Geo-Storage. In Proceedings of the International Petroleum Technology Conference, Dhahran, Saudi Arabia, 12 February 2024. [Google Scholar]
  31. Al-Mukainah, H.; Al-Yaseri, A.; Yekeen, N.; Hamad, J.A.; Mahmoud, M. Wettability of shale–brine–H2 system and H2-brine interfacial tension for assessment of the sealing capacities of shale formations during underground hydrogen storage. Energy Rep. 2022, 8, 8830–8843. [Google Scholar] [CrossRef]
  32. Turkson, J.N.; Md Yusof, M.A.; Darkwah-Owusu, V.; Sokama-Neuyam, Y.A.; Fjelde, I.; Alakbari, F.S.; Tackie-Otoo, B.N. An interpretable machine-learning-driven tool (HyWEC) for hydrogen wettability estimation: Implications for underground hydrogen storage. Int. J. Hydrogen Energy 2025, 155, 150256. [Google Scholar] [CrossRef]
  33. Tackie-Otoo, B.N.; Turkson, J.N.; Darkwah-Owusu, V.; Md Yusof, M.A. Integrating Bayesian Optimization with Machine Learning for Predicting Interfacial Tension in Quaternary Aqueous Hydrogen Systems: Implications for Underground Hydrogen Storage. Arab. J. Sci. Eng. 2025, 50, 21403–21431. [Google Scholar] [CrossRef]
  34. Ishola, N.; Gbadamosi, A.; Muhammed, N.S.; Epelle, E.; Haq, B.; Patil, S.; Al Shehri, D. Data-driven modelling to predict interfacial tension of hydrogen–brine system: Implications for underground hydrogen storage. Results Eng. 2025, 26, 104608. [Google Scholar] [CrossRef]
  35. Thanh, H.V. Revolutionizing hydrogen storage: Predictive modeling of hydrogen-brine interfacial tension using advanced machine learning and optimization technique. Int. J. Hydrogen Energy 2025, 128, 406–424. [Google Scholar] [CrossRef]
  36. Chen, Y.; Tang, S.; Xi, Z.; Sun, S.; Jiao, P.; Mei, X.; Zhou, T.; Sun, J.; Zeng, F. A Machine Learning Approach for Estimating Shale Micro-Mechanical Properties from High Resolution SEM–EDS Images. Rock Mech. Rock Eng. 2025, in press. [Google Scholar] [CrossRef]
  37. Mohanto, S.; Deb, D. Prediction of Plastic Damage Index for Assessing Rib Pillar Stability in Underground Metal Mine Using Multi-variate Regression and Artificial Neural Network Techniques. Geotech. Geol. Eng. 2020, 38, 767–790. [Google Scholar] [CrossRef]
  38. Shin, Y.; Shin, S. Rock Classification in a Vanadiferous Titanomagnetite Deposit Based on Supervised Machine Learning. Minerals 2022, 12, 461. [Google Scholar] [CrossRef]
  39. Honório, H.T.; Houben, M.; Bisdom, K.; van der Linden, A.; de Borst, K.; Sluys, L.J.; Hajibeygi, H. A multi-step calibration strategy for reliable parameter determination of salt rock mechanics constitutive models. Int. J. Rock Mech. Min. Sci. 2024, 183, 105922. [Google Scholar] [CrossRef]
  40. He, Q.; Liu, J.; Wu, F.; Li, C.; Chen, J.; Gao, R.; Ye, C.; Zhu, S. Fatigue properties and constitutive model of Jintan salt rock subjected to complex cyclic loading. Int. J. Min. Sci. Technol. 2025, 35, 121–133. [Google Scholar] [CrossRef]
  41. Guan, W.; Cai, W.; Li, Z.; Lu, H. Microscopic Characterization and Fractal Analysis of Pore Systems for Unconventional Reservoirs. J. Mar. Sci. Eng. 2024, 12, 908. [Google Scholar] [CrossRef]
  42. Zhu, L.; Bijeljic, B.; Blunt, M.J. Diffusion Model-Based Generation of Three-Dimensional Multiphase Pore-Scale Images. Transp. Porous Media 2025, 152, 22. [Google Scholar] [CrossRef]
  43. Seydi, A.; Sadeghnejad, S.; Enzmann, F.; Khoozan, D.; Al-Raoush, R.I.; Kersten, M. Prediction of fine particle retention at the pore-scale using deep learning. Phys. Fluids 2025, 37, 062019. [Google Scholar] [CrossRef]
  44. Maleki, M.; Dehghani, M.R.; Kafi, M.; Akbari, A.; Kazemzadeh, Y.; Ranjbar, A. Machine learning analysis of CO2 and methane adsorption in tight reservoir rocks. Sci. Rep. 2025, 15, 24575. [Google Scholar] [CrossRef]
  45. Kalam, S.; Arif, M.; Raza, A.; Lashari, N.; Mahmoud, M. Data-driven modeling to predict adsorption of hydrogen on shale kerogen: Implication for underground hydrogen storage. Int. J. Coal Geol. 2023, 280, 104386. [Google Scholar] [CrossRef]
  46. Wang, T.; Tang, H. FFINO: Factorized Fourier Improved Neural Operator for Modeling Multiphase Flow in Underground Hydrogen Storage. arXiv 2025, arXiv:2506.17344. [Google Scholar] [CrossRef]
  47. Asghari, M.; Emami Niri, M.; Sedaee, B. UHSNet: Deep learning-based smart proxy modeling for underground hydrogen storage. Energy 2025, 329, 136763. [Google Scholar] [CrossRef]
  48. Sirota, D.D.; Gushchin, K.A.; Khan, S.A.; Kostikov, S.L.; Butov, K.A. Neural Operators for Hydrodynamic Modeling of Underground Gas Storages. Mosc. Univ. Phys. Bull. 2024, 79, S922–S934. [Google Scholar] [CrossRef]
  49. Zhou, J.; Li, Z.; Liu, S.; Li, C.; Zhao, Y.; Zhou, Z.; Liang, G. Hybrid genetic algorithm for parametric optimization of surface pipeline networks in underground natural gas storage harmonized injection and production conditions. Nat. Gas Ind. B 2025, 12, 234–250. [Google Scholar] [CrossRef]
  50. Ma, M.; Zhang, Z.; Chen, J.; Xiong, Q.; Hu, E.; Wang, T.; Xiao, Y. Physics-informed machine learning-driven active utilization of ventilation and heat exchange in large underground tunnels. Renew. Energy 2025, 250, 123283. [Google Scholar] [CrossRef]
  51. Wang, J.; Hu, Z.; Yan, X.; Yao, J.; Sun, H.; Yang, Y.; Zhang, L.; Zhong, J. Gradient-boosted spatiotemporal neural network for simulating underground hydrogen storage in aquifers. J. Comput. Phys. 2025, 521, 113557. [Google Scholar] [CrossRef]
  52. Hosseini, M.; Leonenko, Y. Prediction of hydrogen−brine interfacial tension at subsurface conditions: Implications for hydrogen geo-storage. Int. J. Hydrogen Energy 2024, 58, 485–494. [Google Scholar] [CrossRef]
  53. Hosseini, M.; Leonenko, Y. Predicting interfacial tension in brine-hydrogen/cushion gas systems under subsurface conditions: Implications for hydrogen geo-storage. Int. J. Hydrogen Energy 2024, 91, 1394–1406. [Google Scholar] [CrossRef]
  54. Gbadamosi, A.; Adamu, H.; Usman, J.; Usman, A.G.; Jibril, M.M.; Salami, B.A.; Gbadamosi, S.L.; Oyedele, L.O.; Abba, S.I. New-generation machine learning models as prediction tools for modeling interfacial tension of hydrogen-brine system. Int. J. Hydrogen Energy 2024, 50, 1326–1337. [Google Scholar] [CrossRef]
  55. Xie, M.; Zhang, M.; Jin, Z. Machine Learning-Based Interfacial Tension Equations for (H2 + CO2)-Water/Brine Systems over a Wide Range of Temperature and Pressure. Langmuir 2024, 40, 5369–5377. [Google Scholar] [CrossRef]
  56. Akbari, A. Application of machine learning in modelling gas dispersion coefficients for hydrogen storage in porous media. Can. J. Chem. Eng. 2025. [Google Scholar] [CrossRef]
  57. Vo Thanh, H.; Rahimi, M.; Dai, Z.; Zhang, H.; Zhang, T. Predicting the wettability rocks/minerals-brine-hydrogen system for hydrogen storage: Re-evaluation approach by multi-machine learning scheme. Fuel 2023, 345, 128183. [Google Scholar] [CrossRef]
  58. Lv, Q.; Xue, J.; Li, X.; Rezaei, F.; Larestani, A.; Norouzi-Apourvari, S.; Abdollahi, H.; Hemmati-Sarapardeh, A. White-box machine-learning models for accurate interfacial tension prediction in hydrogen–brine mixtures. Clean Energy 2024, 8, 252–264. [Google Scholar] [CrossRef]
  59. Alqahtani, F.M.; Nait Amar, M.; Djema, H.; Ourabah, K.; Alanazi, A.; Ghasemi, M. Machine Learning–Based Estimation of Hydrogen Solubility in Brine for Underground Storage in Saline Aquifers. Greenh. Gases Sci. Technol. 2025, 15, 409–420. [Google Scholar] [CrossRef]
  60. Huohai, Y.; Qinghui, H.; Chao, M.; Ping, Z.; Hancheng, L.; Yu, F.; Renze, L.; Zonglin, L.; Chenglian, D. Injection-mining scheme optimization of underground gas storage based on agent model. Geoenergy Sci. Eng. 2025, 244, 213406. [Google Scholar] [CrossRef]
  61. Li, Z.; Li, C.; Jia, W.; Chen, Y.; Zhang, C.; Xiao, H.; Pu, Z. Co-optimization method for injection strategy of underground natural gas storage integrating aboveground and underground parts. Gas Sci. Eng. 2024, 128, 205376. [Google Scholar] [CrossRef]
  62. Kuk, E.; Stopa, J.; Kuk, M.; Janiga, D.; Wojnarowski, P. Optimal Well Control Based on Auto-Adaptive Decision Tree—Maximizing Energy Efficiency in High-Nitrogen Underground Gas Storage. Energies 2022, 15, 3413. [Google Scholar] [CrossRef]
  63. Indro, A.; Chellal, H.; Malki, M.; Mao, S.; Gross, M.; Mehana, M. Optimizing underground hydrogen storage performance through multi-well strategies in depleted gas reservoirs. Int. J. Hydrogen Energy 2025, 106, 672–685. [Google Scholar] [CrossRef]
  64. GÜmrah, B.G.F. Gas Production Rate Optimization by Genetic Algorithm. Energy Sources 2001, 23, 295–304. [Google Scholar] [CrossRef]
  65. Guyaguler, B.; Gumrah, F. Comparison of genetic algorithm with linear programming for the optimization of an underground gas-storage field. IN SITU 1999, 23, 131–149. [Google Scholar]
  66. Wamriew, D.; Dorhjie, D.B.; Bogoedov, D.; Pevzner, R.; Maltsev, E.; Charara, M.; Pissarenko, D.; Koroteev, D. Microseismic Monitoring and Analysis Using Cutting-Edge Technology: A Key Enabler for Reservoir Characterization. Remote Sens. 2022, 14, 3417. [Google Scholar] [CrossRef]
  67. Su, F.; Li, X.; Feng, Y.; Li, S.; Wang, Y.; Gu, C.; Si, X. Sophisticated distributed acoustic sensing (DAS) for real-time monitoring and analysis of wellbore integrity. Geoenergy Sci. Eng. 2026, 256, 214162. [Google Scholar] [CrossRef]
  68. Huang, W.; Liao, W.; Li, J.; Qiao, X.; Yusan, S.; Yasen, A.; Li, X.; Zhang, S. The Long-Term Surface Deformation Monitoring and Prediction of Hutubi Gas Storage Reservoir in Xinjiang Based on InSAR and the GWO-VMD-GRU Model. Remote Sens. 2025, 17, 2480. [Google Scholar] [CrossRef]
  69. Zhang, N.; Gao, X.; Lai, X.; Xu, H.; Liu, Y.; Ma, H. Research on crack detection method for shallow-buried underground compressed air energy storage cavern based on improved mask R-CNN model. Earth Energy Sci. 2025, 1, 203–212. [Google Scholar] [CrossRef]
  70. He, Y.; Xu, J.; Wang, M.; Liu, R.; Chen, X.; Sui, J. Intelligent prediction study on the seepage evolution of rock-concrete interface in compressed air energy storage artificial caverns under cyclic loading. Phys. Fluids 2025, 37, 067115. [Google Scholar] [CrossRef]
  71. Gutiérrez-Oribio, D.; Stathas, A.; Stefanou, I. AI-Driven Approach for Sustainable Extraction of Earth’s Subsurface Renewable Energy While Minimizing Seismic Activity. Int. J. Numer. Anal. Methods Geomech. 2025, 49, 1126–1138. [Google Scholar] [CrossRef]
  72. Sugan, M.; Peruzza, L.; Romano, M.A.; Guidarelli, M.; Moratto, L.; Sandron, D.; Plasencia Linares, M.P.; Romanelli, M. Machine learning versus manual earthquake location workflow: Testing LOC-FLOW on an unusually productive microseismic sequence in northeastern Italy. Geomat. Nat. Hazards Risk 2023, 14, 2284120. [Google Scholar] [CrossRef]
  73. Gao, K.; Creasy, N.M.; Huang, L.; Gross, M.R. Underground hydrogen storage leakage detection and characterization based on machine learning of sparse seismic data. Int. J. Hydrogen Energy 2024, 61, 137–161. [Google Scholar] [CrossRef]
  74. Vo Thanh, H.; Safaei-Farouji, M.; Wei, N.; Band, S.S.; Mosavi, A. Knowledge-based rigorous machine learning techniques to predict the deliverability of underground natural gas storage sites for contributing to sustainable development goals. Energy Rep. 2022, 8, 7643–7656. [Google Scholar] [CrossRef]
  75. Wei, W.; Hou, J.; Liu, X. Evaluating and predicting deliverability of natural gas storage sites using stacking machine learning models. Geoenergy Sci. Eng. 2025, 249, 213771. [Google Scholar] [CrossRef]
  76. Derakhshani, R.; Lankof, L.; GhasemiNejad, A.; Zaresefat, M. Artificial intelligence-driven assessment of salt caverns for underground hydrogen storage in Poland. Sci. Rep. 2024, 14, 14246. [Google Scholar] [CrossRef] [PubMed]
  77. Zhao, K.; Yu, S.; Wong, L.N.Y. Long-term stability forecasting for energy storage salt caverns using deep learning-based model. Energy 2025, 319, 134854. [Google Scholar] [CrossRef]
  78. Baronio, E.; Abdo, E.; Mauro, S.; Troise, M.; Salamina, L. Case Study of the Use of a Digital Twin for Leak Detection and Quantification in Underground Gas Storage Wells. In Proceedings of the SPE EuropEC—Europe Energy Conference Featured at the 84th EAGE Annual Conference & Exhibition, Vienna, Austria, 5–8 June 2023. [Google Scholar]
  79. Wang, Z.; Chen, J.; Wang, G.; Li, J.; Li, S.; Azhar, M.U.; Ma, S.; Xu, W.; Zhuang, D.; Zhan, L.; et al. Geometry prediction and design for energy storage salt caverns using artificial neural network. Energy 2024, 308, 132820. [Google Scholar] [CrossRef]
  80. Li, J.; Wang, Z.; Zhang, S.; Shi, X.; Xu, W.; Zhuang, D.; Liu, J.; Li, Q.; Chen, Y. Machine-learning-based capacity prediction and construction parameter optimization for energy storage salt caverns. Energy 2022, 254, 124238. [Google Scholar] [CrossRef]
  81. Li, J.; Liu, J.; Li, K.; Zhang, S.; Xu, W.; Zhuang, D.; Zhan, L.; Chen, Y. Three dimensional interface normal prediction for Volume-of-Fluid method using artificial neural network. Eur. J. Mech. B Fluids 2024, 106, 13–20. [Google Scholar] [CrossRef]
  82. He, T.; Wang, T.; Yang, C.; Liao, Y.; Xie, D.; Ding, Z.; Daemen, J.J.K. A mathematical model for the shape prediction of bedded salt cavern used for energy storage. Int. J. Rock Mech. Min. Sci. 2025, 186, 106029. [Google Scholar] [CrossRef]
  83. Teng, T.; Chen, Y.; Wang, S.; Jia, W.; Wang, Y.; Liu, K.; Li, Z. Water injection softening modeling of hard roof and application in Buertai coal mine. Environ. Earth Sci. 2025, 84, 54. [Google Scholar] [CrossRef]
Energies 18 06354 g001
Figure 1. Main types of UGS (modified after [11,12]).
Figure 1. Main types of UGS (modified after [11,12]).
Energies 18 06354 g001
Energies 18 06354 g002
Figure 2. Procedures for retrieving literature from the WoSCC database.
Figure 2. Procedures for retrieving literature from the WoSCC database.
Energies 18 06354 g002
Energies 18 06354 g003
Figure 3. Trend chart of the number of papers over time.
Figure 3. Trend chart of the number of papers over time.
Energies 18 06354 g003
Energies 18 06354 g004
Figure 4. Network of co-occurring keywords for (a) UGS and (b) AI in UGS.
Figure 4. Network of co-occurring keywords for (a) UGS and (b) AI in UGS.
Energies 18 06354 g004
Energies 18 06354 g005
Figure 5. Research hotspot map of AI in UGS research based on the WOS database.
Figure 5. Research hotspot map of AI in UGS research based on the WOS database.
Energies 18 06354 g005
Energies 18 06354 g006
Figure 6. AI in UGS technology research mutation word statistical chart (top 9 keywords with the strongest citation bursts).
Figure 6. AI in UGS technology research mutation word statistical chart (top 9 keywords with the strongest citation bursts).
Energies 18 06354 g006
Energies 18 06354 g007
Figure 7. XGBoost–SHAP driven fatigue–creep damage modeling and stability assessment of salt cavern gas storage (idea inspired by [40]).
Figure 7. XGBoost–SHAP driven fatigue–creep damage modeling and stability assessment of salt cavern gas storage (idea inspired by [40]).
Energies 18 06354 g007
Energies 18 06354 g008
Figure 8. Gradient-boosted spatiotemporal neural network for simulating underground hydrogen storage in aquifers (idea inspired by [51]).
Figure 8. Gradient-boosted spatiotemporal neural network for simulating underground hydrogen storage in aquifers (idea inspired by [51]).
Energies 18 06354 g008
Energies 18 06354 g009
Figure 9. Accurate IFT prediction in hydrogen–brine mixtures using white-box machine-learning models (idea inspired by [58]).
Figure 9. Accurate IFT prediction in hydrogen–brine mixtures using white-box machine-learning models (idea inspired by [58]).
Energies 18 06354 g009
Energies 18 06354 g010
Figure 10. Enhancing the efficiency of underground hydrogen storage by implementing multi-well development strategies in depleted gas reservoirs (idea inspired by [63]).
Figure 10. Enhancing the efficiency of underground hydrogen storage by implementing multi-well development strategies in depleted gas reservoirs (idea inspired by [63]).
Energies 18 06354 g010
Energies 18 06354 g011
Figure 11. Transformer-based prediction of the long-term thermo-mechanical response of surrounding rock in underground cryogenic storage (idea inspired by [26]).
Figure 11. Transformer-based prediction of the long-term thermo-mechanical response of surrounding rock in underground cryogenic storage (idea inspired by [26]).
Energies 18 06354 g011
Energies 18 06354 g012
Figure 12. Machine-learning-based capacity prediction and construction parameter optimization for UGS salt caverns (idea inspired by [80]).
Figure 12. Machine-learning-based capacity prediction and construction parameter optimization for UGS salt caverns (idea inspired by [80]).
Energies 18 06354 g012
Table 1. Summary of AI tasks, algorithms, data sources, and key challenges for different UGS application scenarios.
Table 1. Summary of AI tasks, algorithms, data sources, and key challenges for different UGS application scenarios.
Application Scenario (Section)AI Task TypeAlgorithmsData SourceKey Challenges/Limitations
3.1 Intelligent characterization of rock mechanical parametersRegression/Prediction/ClassificationXGBoost, ANN, SVM, PSO, GMM, CNN, Autoencoder, U-NetNanoindentation and SEM–EDS tests, XRD composition, micro-CT imaging, pore-scale datasetsLimited high-resolution geomechanical data; sample heterogeneity; model interpretability and scalability
3.2 Multiphysics coupling and surrogate modelingSurrogate modeling/Forecasting/OptimizationFFINO (Fourier Neural Operator), UHSNet, Tensor-decomposed FNO, HGA–GRG, LSTM–Seq2Seq + AttentionMultiphase flow simulations, hydrogen–brine experiments, thermal–fluid monitoring dataCoupling of physics and data; generalization to complex geometries; training cost for large datasets
3.3 Gas–rock–fluid interaction, wettability and interfacial behavior predictionProperty prediction/Regression/InterpretabilityXGBoost, LightGBM, SVR, RBFNN–AGTO, GMDH, GEP, HyWEC (Bayesian Optimization)Molecular simulation results, hydrogen–brine experiments, thermodynamic datasetsData availability and quality; physical interpretability of ML models; transferability to different gas systems
3.4 Injection–production process optimizationMulti-objective optimization/ControlGA, NSGA-III, Reinforcement Learning, Deep Reinforcement Learning, DNN proxy modelsField operation records, numerical reservoir simulations, digital-twin dataNonlinear coupling of surface–subsurface systems; limited field validation; computational cost
3.5 Monitoring, risk assessment and integrity evaluationAnomaly detection/Classification/ForecastingCNN, BiLSTM, CWT–CNN, Mask R-CNN, GWO–VMD–GRU, Transformer, DDPG RLDAS acoustic signals, InSAR data, sensor logs, time-lapse seismic dataData noise and imbalance; real-time processing; model transfer to different storage types
3.6 Intelligent design and construction of underground storage cavernGeometry prediction/Capacity estimation/Design optimizationGRU, BP Neural Network, ANN-based VOF modelMulti-stage leaching records, operational logs, flow simulation dataLack of standardized datasets; limited on-site data for model validation; scaling from lab to field
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Chen, J.; Wang, G.; Bai, X.; Duan, C.; Lu, J.; Xiao, L.; Ge, X.; Zhang, G.; Li, J. Artificial Intelligence for Underground Gas Storage Engineering: A Review with Bibliometric and Knowledge-Graph Insights. Energies 2025, 18, 6354. https://doi.org/10.3390/en18236354

AMA Style

Chen J, Wang G, Bai X, Duan C, Lu J, Xiao L, Ge X, Zhang G, Li J. Artificial Intelligence for Underground Gas Storage Engineering: A Review with Bibliometric and Knowledge-Graph Insights. Energies. 2025; 18(23):6354. https://doi.org/10.3390/en18236354

Chicago/Turabian Style

Chen, Jiasong, Guijiu Wang, Xuefeng Bai, Chong Duan, Jun Lu, Luokun Xiao, Xinbo Ge, Guimin Zhang, and Jinlong Li. 2025. "Artificial Intelligence for Underground Gas Storage Engineering: A Review with Bibliometric and Knowledge-Graph Insights" Energies 18, no. 23: 6354. https://doi.org/10.3390/en18236354

APA Style

Chen, J., Wang, G., Bai, X., Duan, C., Lu, J., Xiao, L., Ge, X., Zhang, G., & Li, J. (2025). Artificial Intelligence for Underground Gas Storage Engineering: A Review with Bibliometric and Knowledge-Graph Insights. Energies, 18(23), 6354. https://doi.org/10.3390/en18236354

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop