1. Introduction
Amid the global energy transition, developing unconventional oil and gas resources has become a key strategy for national energy security [
1,
2]. Subsurface hydraulic fracturing, the core technology behind this unconventional revolution, is shifting from experience-driven approaches to intelligence-driven paradigms [
3]. This transformation is essential for advancing energy technologies and strengthening national energy autonomy.
Traditional hydraulic fracturing, constrained by the linear percolation theory, faces three core technical bottlenecks. First, simplified models based on homogeneous assumptions fail to represent the high heterogeneity of unconventional reservoirs [
4,
5]. Second, static design parameters lack adaptability to complex dynamic subsurface conditions [
6]. Third, experience-driven decision-making lacks both timeliness and accuracy [
7]. These limitations are especially severe in unconventional resource development, restricting improvements in energy self-sufficiency.
To address these technical bottlenecks, artificial intelligence (AI)-driven subsurface hydraulic fracturing has emerged by integrating technologies [
8]. In the perception domain, the industry has moved beyond Euclidean processing [
9]; graph neural networks (e.g., GraphSAGE) now model non-Euclidean heterogeneities with >95% lithology identification accuracy [
10,
11], while the fusion of deep learning with distributed acoustic sensing (DAS; e.g., DASEventNet) has lowered microseismic detection thresholds from -1.14 to -1.80, revealing previously hidden fracture networks [
12]. In cognition and modeling, physics-informed machine learning (PIML) is overcoming black box skepticism [
13,
14]. By embedding conservation laws into neural networks, physics-informed neural networks (PINNs) solve non-Newtonian fluid propagation with over 90% error reduction compared to finite difference methods [
15], and bidirectional long short-term memory (BiLSTM)-attention models now predict proppant dune morphology in real-time with >92% accuracy, replacing prohibitive computational fluid dynamics (CFD)-discrete element method (DEM) simulations [
16,
17]. Crucially, the field is advancing toward autonomous action. Deep reinforcement learning agents (e.g., deep deterministic policy gradient (DDPG)) are transitioning operations from advisory to autonomous control, learning continuous pumping schedules that reduce control errors by over 50% compared to traditional proportional-integral-derivative (PID) or model predictive control (MPC) systems [
18,
19].
However, the transition of these AI models from laboratory validation to field-scale application faces significant hurdles, often described as the lab-to-field gap. Academic models are typically trained on pristine, equilibrated datasets or ideal simulations, which differ fundamentally from the high-noise, heterogeneous, and sparse data environments of actual well sites. Specifically, data heterogeneity resulting from inconsistent sensor standards and variable logging conditions often leads to model failure in deployment, exposing the inadequacy of standalone algorithms to handle complex systemic constraints.
Consequently, achieving industrial scalability requires AI-driven subsurface hydraulic fracturing to be structured around a systematic three-layer architecture: data-driven→dynamic optimization→autonomous decision-making. Through the integration of Internet of Things (IoT), big data analytics, and adaptive control, it enables intelligent management across the entire fracturing process. This approach moves beyond experience-based models and builds a knowledge system encompassing risk identification, parameter optimization, production forecasting, and parameter inversion. At the engineering level, emerging systems function to facilitate multi-source data integration and semi-autonomous decision-making, forming a closed-loop from data perception to decision output. This data- and algorithm-driven paradigm is reshaping operational precision and value creation. As cognitive computing and autonomous control evolve, AI-driven fracturing is progressing from assisted decision-making to self-optimizing systems.
This paper reviews the technical requirements and implementation path of AI-driven fracturing and analyzes its core architecture. It provides theoretical and practical guidance for the efficient development of unconventional oil and gas resources.
2. AI-driven subsurface hydraulic fracturing technology
2.1. Objectives of AI-driven subsurface hydraulic fracturing
AI-driven subsurface hydraulic fracturing aims to develop a next-generation system, as depicted in
Fig. 1, that replaces experience-based decisions with data-driven, scientifically guided operations. By integrating data intelligence, it seeks to improve the efficiency, precision, and sustainability of oil and gas development. The core objective is to build an autonomously evolving system that enhances the economic performance and long-term viability of unconventional oil and gas resource development.
At the technical level, AI-driven fracturing targets three key breakthroughs: ① full-process autonomous optimization, where systems perceive formation responses in real time and dynamically adjust operational parameters to match geological conditions; ② reservoir-scale collaborative development, enabling multi-well coordination and intelligent block-level planning to maximize reservoir drainage; ③ continuous system evolution, by integrating emerging technologies such as digital twins and edge computing to support self-learning and iterative optimization.
At the social value level, AI-driven fracturing focuses on three core objectives: zero-waste resource utilization by precisely controlling material usage to achieve full efficiency of fracturing inputs; zero-harm environmental impact by minimizing risks to groundwater and surface ecosystems; low-carbon energy production by improving operational efficiency and reducing carbon emissions per unit output.
As a key enabler of smart oilfield development, AI-driven fracturing is propelling the digital transformation of traditional oil and gas operations while supporting long-term energy sustainability. Its industrial feasibility has been demonstrated through multiple successful applications: Halliburton’s Prodigi® system enables real-time control of pumping rates, Schlumberger’s Kinetix® platform optimizes fracturing parameters using integrated multi-source data, and Baker Hughes’ XACT® system delivers millisecond-level downhole monitoring via acoustic analysis. These advancements lay a solid foundation for scaling AI-driven fracturing across the industry and accelerating the digital evolution of oilfield operations.
2.2. Architecture of AI-driven subsurface hydraulic fracturing
AI- driven subsurface hydraulic fracturing is redefining the development model for unconventional oil and gas. It establishes a multi-modal perception system that integrates surface, downhole, and spatial sensing; builds hybrid modeling frameworks combining physical mechanisms with data-driven methods; and forms a closed-loop control system for autonomous decision-making. This enables end-to-end intelligence—from microscopic fracture propagation to field-scale engineering execution. The shift from experience-based to cognition-driven approaches addresses the efficiency limitations of conventional fracturing and marks a new phase in digital energy development.
To realize this vision, AI-driven fracturing follows a three-principle framework: data-driven, dynamic optimization, and autonomous decision-making. This forms a systematic technology architecture through multidisciplinary integration. As shown in
Fig. 2, this framework consists of four hierarchical layers.
(1) Data perception, transmission, and integration layer. This foundational layer establishes a multi-dimensional data acquisition system. High-precision sensors and real-time monitoring devices enable comprehensive sensing of formation properties, fracturing fluid behavior, and fracture propagation dynamics [
20,
21]. Using anti-interference wireless protocols and fiber-optic hybrid transmission [
22], it builds reliable, high-speed downhole-to-surface channels. IoT integration further consolidates multi-source heterogeneous data into a unified platform, enabling downstream analysis and control.
(2) Intelligent analysis and modeling layer. Building on integrated data, this layer applies AI and machine learning to develop dynamic response models. It integrates physics-based models with real-time data to accurately predict fracture geometry and stress evolution [
23], supporting optimization decisions.
(3) Dynamic optimization and control layer. Driven by model outputs, this layer dynamically adjusts key fracturing parameters—such as flow rate, pressure, and fluid viscosity—to adapt to subsurface heterogeneity. It aims to build a closed-loop feedback system to keep operations in optimal conditions [
24].
(4) Autonomous decision-making and execution layer. As the top layer, this stage aims to enable unmanned or minimally manned fracturing operations. By combining expert knowledge with algorithms, the system can autonomously detect anomalies and respond in real time [
25]. It also features long-term learning capabilities to continuously improve operational efficiency and safety.
Critically, the synergy among these four layers is governed by a rigorous bidirectional data-instruction transmission mechanism and real-time interaction protocols. The architecture operates through an upward abstraction flow, where raw heterogeneous signals acquired by the perception layer are transformed into interpretable physical descriptors in the modeling layer, eventually culminating in strategic logic within the autonomous decision layer. Conversely, a downward command cascading flow translates high-level directives into specific parameter setpoints executed by actuators at the foundational layer. This hierarchy is sustained by dynamic inter-layer feedback loops, where model uncertainties or operational deviations trigger adaptive high-frequency sampling or real-time model calibration, ensuring a synchronized, self-correcting closed-loop ecosystem.
This framework transforms traditional fracturing operations in three key ways: shifting from static design to real-time dynamic optimization, from human-driven decisions to autonomous algorithmic control, and from isolated processes to full-process intelligent collaboration. By building a closed-loop system of perception-analysis-optimization-execution, AI-driven fracturing significantly enhances operational efficiency and safety. It drives the industry into a new stage of self-adaptive, self-learning, and self-optimizing development. Current AI-driven fracturing design goes beyond traditional empirical models by using deep learning integrated with reservoir simulation and geomechanical constraints. It builds proxy models for collaborative optimization, enabling multi-objective dynamic design—maximizing stimulated reservoir volume and ensuring balanced fracture growth. This approach greatly improves stimulation performance and recovery efficiency. The implementation of this framework offers a scalable model for the digital transformation of oil and gas development, with potential applications in other energy sectors.
3. Preliminary practices in AI-driven subsurface hydraulic fracturing
In recent years, a full-chain technical system for AI-driven hydraulic fracturing—from fracture prediction to production optimization—has been established and validated in practical applications.
3.1. AI-based prediction of 3D fracture propagation
To efficiently predict dynamic fracture propagation in shale reservoirs, a deep learning model named Dy-Fracture-Net can be developed [
26], based on a parallel semantic segmentation architecture, as illustrated in
Fig. 3. The model integrates multi-source heterogeneous data through a three-part framework:
Stratified feature extraction. Parallel convolutional layers aligned with reservoir strata extract spatially relevant features from formation data.
Pumping feature enhancement. Fast Fourier transform (FFT), convolutional operations, and residual connections are applied to capture complex nonlinear relationships between pumping parameters and fracture morphology.
Multi-source data fusion. Learnable channel attention weights dynamically highlight critical features from different data sources, improving prediction accuracy.
The core of the model is a 3D spatiotemporal semantic segmentation network that integrates multi-source heterogeneous data. In the encoding stage, batch normalization accelerates convergence by standardizing data distributions, while multi-stage convolutional blocks extract and refine key features. During decoding, the model combines low- and high-level features via upsampling, convolution, and concatenation, followed by Sigmoid activation and implicit thresholding to achieve pixel-level identification of fracture zones.
This architecture captures reservoir heterogeneity and spatiotemporal dynamics, forming a complete segmentation pathway from data encoding to semantic prediction. By effectively fusing pumping parameters, reservoir properties, and spatial structures, the model enables accurate prediction of dynamic 3D fracture propagation. As evidenced in
Fig. 4, while minor deviations persist at the fracture boundaries, the topology of the main fractures, secondary fracture networks, and fracture height containment across different layers are predicted with high fidelity. Quantitative evaluation confirms an overall fitting accuracy exceeding 90%. Critically, in terms of computational efficiency, the model achieves a dramatic acceleration, reducing the simulation time per stage from approximately 30 min (typical of conventional numerical methods) to just 10 s, thereby satisfying the latency requirements for real-time field decision-making.
3.2. AI-driven early warning and dynamic control of fracturing processes
To tackle challenges such as low fracturing efficiency, limited accuracy of manual identification, and delayed safety responses in shale gas wells, this study proposes an integrated solution combining intelligent data processing, multi-model analysis, and real-time optimization control, as illustrated in the real-time data-driven fracturing parameter optimization system shown in
Fig. 5.
In the data preprocessing stage, a mean filter with a kernel size of 5 is used to remove noise from raw time-series signals. Statistical features are extracted, and sliding window sampling with a window size of 300 time steps and a step length of 1 is applied to structure the data for modeling. The core modeling framework adopts a dual-model collaborative architecture.
The first model is a hybrid neural network combining multi-layer perceptron (MLP)-BiLSTM, where the MLP module performs deep nonlinear feature extraction and the BiLSTM captures bidirectional temporal dependencies. This model enables accurate identification of key fracturing events such as formation breakdown, sudden pump shutdown, and sand blockage. Field validation confirms its precision, showing pressure identification deviations of less than 0.3 MPa and temporal latencies within 15 s for breakdown and shut-in events; even for complex anomalies like sand plugging, event boundary recognition errors are controlled within 25 s, ensuring timely safety alerts.
The second model addresses pressure prediction. It integrates convolutional layers for local pattern recognition, attention mechanisms for key feature focusing, BiLSTM for temporal dynamics, and skip connections for stable gradient flow. This structure achieves high-accuracy pressure forecasting over a 120 s time horizon. Quantitative assessment against actual post-fracturing data reveals a fitting accuracy exceeding 95% for this rolling forecast window, providing a robust and reliable basis for preemptive control adjustments.
These models function through a predictive-input cascade mechanism: The 120 s pressure forecast generated by the prediction model is directly fed into the recognition model. This coupling allows the system to identify latent risks, such as sand plugging [
25], on projected trajectories, thereby enabling preemptive warnings and shifting operations from reactive detection to proactive control.
Following model construction, a two-stage training strategy with transfer learning is employed to improve generalization in small-sample conditions (fewer than 20 labeled samples per category). Validation confirms that this approach maintains an event recognition accuracy exceeding 90% despite data scarcity. Based on model outputs, a closed-loop control system is established to integrate event detection, risk evaluation, and parameter optimization, as illustrated in
Fig. 6, which shows the real-time identification and adaptive control mechanism for proppant plugging risks. By tracking real-time changes in key variables, the system dynamically adjusts operating parameters. This real-time optimization process is demonstrated in
Fig. 7, where pressure distribution before and after optimization is compared—
Fig. 7(a) shows abnormal pressure behavior prior to adjustment, while
Fig. 7(b) illustrates the improved control outcome following optimization.
3.3. AI-based synchronized prediction of post-fracturing production and reservoir dynamics
To overcome the technical challenges of forecasting post-fracturing production behavior in shale gas wells, this study introduces Dy-Production-Net, a multi-module fusion prediction network based on 3D residual spatiotemporal convolutions [
27]. By combining hierarchical feature fusion and spatiotemporal dependency modeling, the network enables integrated, high-resolution prediction of reservoir pressure, saturation, and gas production.
The model adopts a hierarchical modular architecture consisting of three key components: a 3D feature fusion module, a dual-channel dynamic output module, and a residual production prediction module, as illustrated in
Fig. 8. The fusion module uses 3D convolutional kernels to extract spatial features of reservoir properties, while 3D max pooling reduces dimensionality. Positional embedding enhances spatial awareness, and multi-head self-attention captures inter-attribute dependencies. The output module follows a U-Net-style encoder-decoder structure, where 3D linear interpolation upsampling feature maps, and skip connections integrate multi-scale features. It outputs predicted pressure and saturation fields through dedicated convolutional branches.
For validation, a representative shale reservoir model was selected, characterized by a porosity range of 4%-10% and an ultra-low matrix permeability ranging from 0.001 to 0.002 mD (1 mD = 10
-16 m
2). The model’s effectiveness in this regard is shown in
Fig. 9, where the post-fracturing pressure and saturation field predictions demonstrate strong consistency with high-fidelity numerical simulation benchmarks; specifically, quantitative error analysis reveals that the mean absolute error (MAE) for the pressure field is maintained below 0.065 MPa, while the saturation field error is controlled within 0.85%. The production prediction module comprises residual blocks with paired 3D convolutional layers to address vanishing gradients, followed by fully connected layers to regress final gas production. A comparison between model predictions and actual field measurements of daily and cumulative gas production is presented in
Fig. 10, where a fitting accuracy exceeding 97% is achieved for both trajectories, forcefully illustrating the model’s accuracy and robustness.
This model introduces a spatiotemporal encoding method that transforms production time-series data into channel dimensions, establishing explicit temporal-spatial correlations. A multi-level feature fusion mechanism based on a feature pyramid structure integrates representations across abstraction levels, while an adaptive feature weighting module dynamically adjusts the contribution of each channel.
Leveraging a multi-module fusion strategy, the architecture enables coordinated interaction among specialized components. During training, a hybrid strategy with sub-module partitioning mitigates convergence issues caused by model complexity, target diversity, and dimensional inconsistency through weight freezing and staged training. Additionally, a two-step optimization process assisted by transfer learning accelerates convergence and enhances model generalization.
3.4. AI-driven real-time optimization of fracturing operations
To achieve dual optimization of production enhancement and economic efficiency in shale gas fracturing, this study develops an AI-driven real-time decision-making framework for real-time pumping program optimization. Built upon previously established technologies—including Dy-Fracture-Net for 3D fracture propagation prediction, intelligent early warning and real-time control systems, and Dy-Production-Net for post-fracturing production and reservoir dynamics forecasting—the approach focuses on three key pumping parameters: pressure, slurry rate, and sand concentration. By integrating multi-source data with predictive models, it constructs a closed-loop prediction-evaluation-optimization-control decision architecture for real-time operational guidance.
The system’s core philosophy integrates data-driven models with physics-based constraints to establish an adaptive optimization framework. It leverages deep predictive models—Dy-Fracture-Net and Dy-Production-Net—to provide real-time insights into fracture propagation, reservoir stress changes, and future production trends. At the same time, event detection and pressure prediction modules identify anomalies such as instantaneous pump shutdowns, proppant bridging, and overpressure events, enabling rapid risk assessment and timely response strategies. This integrated approach is illustrated in
Fig. 11, which outlines the full-process workflow of modeling and optimization for hydraulic fracturing.
For algorithm design, the study introduces multi-objective reinforcement learning (MORL) to manage field complexity and conflicting operational goals. The approach defines three core objectives—maximizing production, improving fracturing efficiency, and minimizing safety risks—while incorporating geomechanical models, empirical rules, and historical operational data to build an intelligent optimization network with self-adaptive policy capabilities. Through continuous interaction with the environment and reward-based learning, the policy network autonomously identifies optimal pumping parameter trajectories, demonstrating strong generalization in complex nonlinear scenarios, such as high-stress-difference formations and multi-branch fracture competition.
During field operations, the system employs a real-time, feedback-driven closed-loop control mechanism. High-frequency acquisition of pressure, displacement, and sand concentration data feeds into proxy models and the MORL module, enabling rapid response and continuous optimization. When discrepancies arise between actual performance and model predictions, the system automatically re-calibrates its optimization strategy, dynamically adjusting pumping parameters—such as stage length, injection rate, and sand ratio—to ensure the fracturing process remains within optimal control boundaries. As illustrated in
Fig. 12, the system effectively balances multiple competing objectives—such as production maximization, operational risk mitigation, and fracturing efficiency—and produces optimized operational trajectories that outperform traditional single-objective approaches.
4. Future prospects for AI-driven subsurface hydraulic fracturing
Current AI-driven fracturing technologies are evolving from reliance on limited surface data toward integrated downhole multi-source data synergy. Despite notable advancements, several key challenges remain:
(1) Harsh downhole environments hinder multi-source data acquisition and compatibility, with real-time transmission vulnerable to interference; dedicated chips and efficient algorithms for real-time processing are still lacking.
(2) Purely data-driven models struggle to incorporate multiphase flow physics, resulting in limited interpretability and poor generalizability across geological scenarios.
(3) Intelligent optimization algorithms face difficulties in multi-objective coordination and adapting to complex reservoir conditions, while lacking autonomous closed-loop control capabilities for real-time decision-making and regulation.
As depicted in
Fig. 13, the target architecture of next-generation AI-driven fracturing systems envisions deep integration of multi-modal sensing, intelligent modeling, and autonomous decision-making. In line with this vision, future breakthroughs must prioritize three directions: miniaturizing multi-modal sensing agents, developing self-interpretable mechanism-data fusion models, and advancing closed-loop control from human-in-the-loop to fully autonomous decision-making—ultimately achieving a qualitative leap from experience-driven to cognition-driven fracturing.
4.1. Multi-modal data sensing agents
As fracturing operations advance into deeper, more complex formations, downhole environments exhibit intensified heterogeneity, stress sensitivity, and dynamic fracture evolution. Developing multi-modal sensing agents, as illustrated in
Fig. 14, with real-time sensing, autonomous communication, and dynamic integration capabilities becomes critical for efficient fracturing control.
Real-time data perception capability. DAS/distributed temperature sensing (DTS) and micro-electromechanical systems (MEMS) enable high-resolution monitoring of wellbore strain, temperature, and microseismic events, demonstrating the efficacy of integrated multi-parameter sensors for fracture diagnostics [
28]. However, traditional electronic sensors frequently suffer from drift and failure under high-pressure, high-temperature (HPHT) and corrosive conditions, while current technologies struggle to synchronously capture the multi-physics coupling signals (e.g., electromagnetic-acoustic responses) required to fully characterize complex fracture networks. To serve as the reliable nerve endings of intelligent systems, future sensing agents must leverage advanced materials to ensure survival in extreme environments and evolve toward synchronized multi-parameter monitoring [
29]. This evolution facilitates the accurate capture of both conventional engineering metrics and transient fracture propagation signals, establishing a robust, high-fidelity data foundation for intelligent decision-making.
Autonomous data communication capability. Wireless sensor networks (WSN) attempt to address the connectivity limitations of standard mud-pulse and acoustic telemetry but remain constrained by low bandwidth, high formation attenuation, and interference [
30,
31]. Such physical bottlenecks result in data islands, preventing the real-time collaboration essential for swarm intelligence [
32]. Consequently, the next generation of subsurface systems requires robust, self-organizing communication networks [
33]. These systems, integrating anti-interference transmission and inter-agent protocols, function to enable rapid local data exchange. This connectivity empowers agents to dynamically optimize perception strategies and sampling frequencies in response to environmental feedback, establishing a foundation for autonomous closed-loop control.
Dynamic data integration capability. Centralized processing paradigms, which transmit raw data to surface servers, create unacceptable latency and bandwidth constraints, often precluding real-time control during high-frequency fracturing operations. The sheer volume of redundant data further impedes operational efficiency. Overcoming these limitations requires the deployment of sensing agents with edge-enabled dynamic integration capabilities [
34]. Through local preprocessing tasks such as feature extraction and preliminary fusion, these agents filter redundancy at the source. This shift to edge intelligence enables the on-demand activation of analysis modules, ensuring the seamless integration of high-value data with surface control systems [
35].
4.2. Multi-parameter synergistic AI-driven fracturing modeling
Accurate simulation of fracturing processes and lifecycle optimization require advanced multi-parameter synergistic modeling, as outlined in
Fig. 15. Serving as the core brain of AI-driven fracturing, this modeling integrates reservoir geological modeling, physics-informed deep learning, and digital twin technology to establish a comprehensive, accurate, and predictive framework.
AI-enhanced reservoir geological modeling. Machine learning has successfully accelerated the processing of seismic and logging data, yet traditional geological models remain largely static, failing to dynamically reflect the spatiotemporal variations in reservoir properties induced by fracturing stress changes. This static nature creates information silos where real-time downhole dynamics are not fed back into the geological understanding. Future intelligent geological modeling must evolve into a dynamic information hub that eliminates data silos by integrating multi-source data (downhole dynamics, lab analysis, and remote sensing) to construct self-evolving probabilistic models [
35]; equipped with knowledge reasoning capabilities, these models will proactively interact with engineers and continuously update through feedback-driven learning, serving as the core geological intelligence of the system [
36,
37].
Data-mechanism dual-driven modeling. Pure data-driven models are inherently limited by their prohibitive data requirements and lack of physical interpretability. Although PINNs offer a dual-driven alternative, they suffer from optimization instability in phase-field fracture modeling, primarily due to the competing gradient scales between sharp fracture interfaces and the bulk continuum, which leads to convergence failure [
13,
38,
39]. To overcome these bottlenecks, future dual-driven modeling must evolve from penalty-based soft constraints to architecture-embedded hard constraints. By integrating physical conservation laws directly into the neural network’s topological design and employing operator-theoretic approaches, next-generation models will achieve strict physical consistency and robust adaptability in complex, dynamic fracturing environments [
40].
Digital twin technology applications. Existing digital twins often suffer from significant latency and unidirectional data flow, rendering them incapable of synchronizing with the millisecond-scale transient changes characteristic of hydraulic fracturing [
41]. This limitation confines them to the role of static observers, unable to support dynamic decision-making under uncertainty. The next generation of digital twins must evolve into active predictive agents. By integrating comprehensive multi-physics mapping, these systems will support real-time counterfactual analysis, empowering engineers to simulate and stress-test various fracturing parameters in a risk-free virtual environment [
42,
43]. This shift from monitoring to preemptive control ensures that operational effectiveness is maximized and risks are mitigated before any physical intervention occurs.
4.3. Closed-loop real-time optimization and control system
To bridge AI-driven fracturing from theoretical research to industrial-scale deployment, developing a closed-loop real-time optimization control system is imperative. The system’s overall framework is shown in
Fig. 16, where it acts as the command center, integrating strategy generation, instruction execution, feedback loops, and solution optimization to deliver precise and efficient fracturing operations.
Optimization control computing architecture. Reliance on manual setpoints and PID loops renders current systems incapable of processing multi-source heterogeneous data or managing conflicting operational objectives in real-time. The resulting computational latency prevents effective responses to millisecond-level pressure fluctuations. To address these deficits, future optimization control architectures must be built upon ultra-high-speed parallel computing frameworks. These systems require self-evolving logic that autonomously refines control strategies based on accumulated data. This capability ensures the rapid interpretation of sensor inputs and the robust, dynamic optimization of operations, adapting seamlessly to complex subsurface conditions.
Dual closed-loop feedback optimization. Existing semi-automated systems lack the adaptive capacity to calibrate internal models against evolving reservoir conditions, causing unavoidable performance drift. Future autonomous systems must therefore rely on a hierarchical monitoring-modeling-decision-verification framework underpinned by a dual closed-loop feedback mechanism. This architecture integrates a primary optimization loop for millisecond-level process control with a parameter calibration loop for dynamic model updating. Such a dual-driven approach harmonizes instantaneous optimization with evolutionary refinement, ensuring that immediate control precision is maintained alongside long-term predictive accuracy.
System reliability and security. The escalation of autonomy and connectivity in fracturing systems amplifies exposure to component failures and cyber-physical threats, creating vulnerabilities that current implementations often overlook. Eliminating single points of failure (SPOF) and insecure protocols is critical to preventing catastrophic operational disruption. Future architectures must therefore adopt a multilayered defense-in-depth framework: enforcing hardware redundancy, integrating software-based fault tolerance and self-diagnostics, and deploying rigorous cybersecurity protocols (encryption and firewalls). This holistic approach ensures the system’s survival and data integrity within high-risk, extreme engineering environments.
The sustained evolution of AI-driven subsurface hydraulic fracturing is not only fundamental to the efficient development of unconventional resources but also integral to the construction of smart oilfields. Future innovations in advanced sensing, high-performance simulation platforms, deep learning, and digital twin technologies will drive AI-driven fracturing toward greater system integration, autonomy, and scalability—ultimately serving as a key enabler for enhancing energy security and optimizing resource development efficiency.
5. Conclusions
This paper presents a comprehensive review of the current development, key technological advances, and future directions of AI-driven fracturing, summarizing a technical evolution framework transitioning from data-driven analysis to dynamic optimization and ultimately to autonomous decision-making. The research demonstrates that AI-driven fracturing is catalyzing a paradigm shift in hydrocarbon development—from traditional experience-based methods to intelligent, data-centric approaches. The proposed Dy-Fracture-Net model enables high-precision prediction of three-dimensional fracture propagation. The dual-model collaborative architecture facilitates intelligent early warning and dynamic optimization of fracturing operations. Meanwhile, the Dy-Production-Net network integrates the prediction of post-fracturing reservoir parameters and production performance. By incorporating intelligent optimization algorithms, a real-time closed-loop control system is established—covering the full lifecycle from fracturing to production and enabling intelligent, adaptive regulation.
Technological prospects point to three key directions for future breakthroughs: ① miniaturization and enhanced intelligence of multi-modal sensing agents; ② development of self-interpretable models that integrate physical mechanisms and data-driven learning; and ③ transition from human-in-the-loop to fully autonomous decision-making. These advances will catalyze a fundamental shift from experience-based to cognition-driven fracturing operations.
This study contributes threefold across theoretical, methodological, and practical dimensions. ① Theoretical: establishes a systematic framework for AI-driven fracturing technology; ② methodological: introduces new paradigms for multi-source data fusion and hybrid modeling; ③ practical: proposes technically feasible solutions for efficient unconventional resource development.
The widespread adoption of AI-driven fracturing is expected to significantly improve hydrocarbon recovery rates and reduce development costs—playing a vital role in ensuring national energy security and advancing dual-carbon objectives. Future research should prioritize breakthroughs in miniaturized sensors, explainable AI algorithms, and autonomous decision-making systems to accelerate industrial-scale deployment.
PDF
(5759KB)
