Development Efficiency Assessment of Challenging Hydrates Under Reservoir Fracturing and Thermal Stimulation Using an XGBoost-SHAP Framework

Abstract

Reservoir fracturing combined with thermal stimulation is a highly promising strategy for the development of challenging hydrates. However, the synergistic influence mechanisms of multiple engineering parameters on productivity remain poorly understood. In this study, based on the geological condition of the SH2 site in the Shenhu Area of the South China Sea, a numerical model was built to investigate the development efficiency of challenging hydrates under fracturing and thermal co-stimulation. Using average gas production rates (m³/d) at recovery rates of 0.70 and 0.85 as assessment indicators, eXtreme Gradient Boosting (XGBoost) and SHapley Additive exPlanations (SHAP) algorithms were employed to quantitatively measure multivariable importance. The results indicated that enhancing the inter-well interaction through reservoir fracturing can increase development efficiency by 2 to 5 times; however, it is not the case that larger-scale fracturing is always preferable, as it can lead to more severe water flooding. Additionally, data-driven models revealed that fracture length (SHAP values of 15.55 and 9.19) was the primary factor influencing development efficiency, followed by the fracture conductivity (SHAP values of 6.65 and 6.32), whereas injection pressure (SHAP values of 2.90 and 2.17), injection temperature (SHAP values of 2.41 and 2.13), and production pressure (SHAP values of 2.37 and 1.82) had relatively limited influences. Most importantly, the positive interaction effect between fracture length and fracture conductivity cannot be ignored. In our simulation, the recommended fracture length and conductivity were 40 m and 100 D·cm, respectively. These findings provide important insights and guidance for implementing this novel co-stimulation method in challenging hydrates.

1. Introduction

Natural gas hydrate is a solid, ice-like crystalline substance formed when guest gas molecules—primarily methane—become trapped within hydrogen-bonded water cavities under low-temperature and high-pressure environments. In nature, methane hydrate (MH) is widely distributed in marine continental margin sediments and terrestrial permafrost regions. Owing to the enormous global reserves (containing 2.5–120 × 10¹⁵ STm³ of natural gas [1,2]) and clean combustion characteristics, MH has attracted considerable attention as a potential transitional energy source that could significantly contribute to future energy security and climate change mitigation.

Since the strategic importance of MH was recognized in the 1980s, a global boom in hydrate exploration and development has taken hold. Nevertheless, unique resource characteristics—such as deep-water drilling complexities, peculiar phase transitions, wellbore instability, multiphase flow, sand production, gas leakage, submarine slope failure, and low productivity—have severely limited its commercialization process [3,4,5]. Furthermore, the vast majority of deposits exhibit distinctive characteristics, including significant thermodynamic stability, low in situ temperatures, limited intrinsic permeability, and the absence of impermeable boundaries. These characteristics collectively define the category of so-called “challenging hydrates,” making extraction particularly challenging [6]. A representative case is the deposits in the Shenhu Area of the South China Sea, which consist of low-permeability bounded by permeable boundary layers, a configuration typical of this challenging category. Despite these unfavorable conditions, two recent pilot production trials conducted have demonstrated the technical feasibility of developing such challenging deposits, marking a notable advancement in the field [7,8].

The combination of depressurization and thermal stimulation (particularly those involving hot fluid injection) is widely regarded as a particularly promising hydrate extraction approach due to its capacity to supply the heat required for sustained hydrate dissociation [9,10,11,12]. Particularly, multi-well injection-production modes have demonstrated significant potential by improving mass and heat transfer efficiency, and have accordingly been suggested as a candidate strategy for commercial-scale hydrate development, notably in the Nankai Trough of Japan [13,14,15]. However, the direct transfer of this development mode to Shenhu hydrates may encounter significant obstacles. The fundamental difference lies in the fact that the permeability of sandy deposits in the Nankai Trough is 1 to 3 orders of magnitude higher than that of fine-grained silty deposits in the Shenhu Area. Such low permeability is extremely detrimental to inter-well mass and heat transfer. For instance, thermal stimulation effects are often confined to localized areas [16], and insufficient inter-well connectivity prevents the recovery of gas generated by hydrate thermal decomposition [13,17]. Furthermore, the absence of continuous impermeable boundaries leads to significant heat loss to adjacent formations, thereby reducing energy utilization efficiency [18]. Perhaps most critically, high injection pressures may drive gas upward, thereby elevating geo-environmental risks such as methane leakage [19,20,21]. Addressing these interrelated challenges is therefore essential for the application of injection-production modes in challenging hydrates.

Reservoir fracturing has been extensively employed as a key enabling technology for the development of low-permeability oil and gas resources. In light of its success in those applications, it has been proposed to enhance the recovery efficiency of challenging hydrates [22,23,24]. Although hydrate deposits are typically weakly consolidated and non-lithified, which presents considerable difficulties for fracturing operations—such as high initiation pressure, the need for high-viscosity fracturing fluids—a growing body of research, including our previous studies regarding fracability evaluation—has confirmed the feasibility of inducing fractures in such complex geomaterials [25,26,27,28,29]. Furthermore, the stimulation capability of the fractures has also been preliminarily demonstrated. Chen et al. [30] and Yang et al. [31] found that fractures can accelerate low-pressure propagation, reduce fluid flow resistance, thereby significantly improving both hydrate dissociation efficiency and gas production rates. Sun et al. [32] reported that horizontal fractures outperform vertical fractures in enhancing productivity. However, after 400 days of production, the stimulative effect diminishes due to insufficient reservoir sensible heat, a trend also observed by Guo et al. [33]. Li et al. [34] further proposed an optimal fracturing design featuring 4 to 6 horizontal fractures with a width of 5 cm, radius of 40 m, and permeability of 1 × 10⁻¹⁰ m². Shen et al. [35] demonstrated that in permafrost-associated hydrate reservoirs, fracturing effectively mitigates the adverse effects of ice formation on gas production. Ju et al. [36] showed that fracturing enlarges the zone influenced by hot water injection, thereby promoting more extensive hydrate dissociation. Similarly, Zhong et al. [37,38], Ma et al. [39], and Nie et al. [40] reported that fracturing substantially enhances inter-well mass and heat transfer, and suggested that such improvements could enable commercially viable gas extraction from challenging hydrate. The aforementioned studies reveal that reservoir fracturing can greatly enhance the injection-production development potential of challenging hydrates. However, the ultimate effectiveness of this novel complex plan hinges upon multiple engineering factors, particularly fracture parameters and injection-production conditions. Furthermore, a systematic assessment of how these coupled parameters influence gas recovery efficiency remains lacking, leaving a gap in the reference framework for implementing engineering solutions.

This study aims to evaluate the development efficiency of challenging hydrates under the co-stimulation of fracturing and thermal, comprehensively considering multiple key engineering parameters. The unique contributions of this work include: (1) systematically investigating the production performance under the coupled effect of reservoir fracturing and thermal stimulation; (2) quantitatively ranking the importance of key parameters and identifying their synergistic effects using an interpretable machine learning approach, which offers new insights and practical guidance for the design and optimization of stimulation strategies in challenging hydrates.

2. Modeling

2.1. Conceptual Model

Figure 1 shows a schematic diagram of the proposed development plan for deep-sea hydrates. A horizontal well pattern is employed, with all horizontal sections located in the middle of the hydrate-bearing layer (HBL). The injection and production wells were arranged in an alternating configuration, and horizontal well fracturing technology was adopted to enhance inter-well connectivity, a widely used reservoir stimulation technique for low-permeability oil and gas reservoirs.

In this study, the challenging hydrate deposits at the SH2 site in the Shenhu Area of the South China Sea served as a typical case study, and the potential of this novel development plan was systematically evaluated. Additionally, assuming that the hydrate deposit is homogeneous along the wellbore axis and considering the symmetry of the development plan configuration, the simulation is based solely on a unit horizontal section length and a single injection-production unit.

2.2. Numerical Code

To investigate the injection-production behavior in fractured hydrate deposits, a modeling approach that incorporates hydrate reaction kinetics or thermodynamic equilibrium, mass and energy conservation, and multiphase fluid flow is required. The Tough + Hydrate simulator, a widely used numerical code in hydrate-related studies, provides a comprehensive framework for such simulations [41]. In this code, a finite-difference integration method is used in the spatial domain, and a fully implicit difference scheme—specifically, a forward-differencing scheme—is used in the temporal domain, which has the advantage of unconditional convergence. The Newton–Raphson method is employed for iterative solution. To ensure solution accuracy, the initial time step was set to 1 s, with a maximum of 10 days. In this model, reservoir fractures were conceptualized as porous media with enhanced permeability, and multiphase fluid flow within them is governed by Darcy’s law [42]. Additionally, its reliability has been rigorously validated through laboratory experiments and field-scale applications [43,44], supporting its suitability for the present study. Accordingly, this simulator is employed to conduct the numerical investigations presented herein.

2.3. Initial and Boundary Conditions

The initial reservoir temperature was reported as 14.87 °C, and a geothermal gradient of 0.047 °C/m was applied, allowing the temperature distribution throughout the domain to be determined [45]. The pressure field is derived from a reference reservoir pressure of 14.97 MPa and a hydrostatic pressure gradient [46]. Given the sufficient thickness of the overlayer and underlayer, constant temperature and pressure boundary conditions were applied to the top and bottom of the model. Detailed model parameters are provided in

Table 1. Model parameters.

Parameters	Value
Grain density	2600 kg/m3
Dry and wet thermal conductivity of deposits	1 and 3.1 W/m/K
Grain-specific heat	1000 J·K/kg
Porewater salinity	0.03
Formation thickness (Overlayer, HBL, and underlayer)	80, 40, 80 m
Wellbore diameter	0.2 m
Well spacing	100 m
Intrinsic permeability and porosity of wells	5 × 10−9 m2 and 1 [6]
Intrinsic permeability and porosity of deposits	10 mD and 0.38
Aqueous saturation (Overlayer, HBL, and underlayer) ${\mathrm{S}}_{\mathrm{A}}$	1, 0.56, and 1
Hydrate saturation ${\mathrm{S}}_{\mathrm{H}}$	0.44
Calculation model for capillary pressure ${\mathrm{P}}_{\mathrm{c}\mathrm{G}\mathrm{W}}$	${\mathrm{P}}_{\mathrm{c}\mathrm{G}\mathrm{W}}=-{\mathrm{P}}_0{[{\left({\displaystyle \frac{{\mathrm{S}}_{\mathrm{A}}-{\mathrm{S}}_{\mathrm{i}\mathrm{r}\mathrm{A}}}{1-{\mathrm{S}}_{\mathrm{i}\mathrm{r}\mathrm{A}}}}\right)}^{-1/\mathsf{\vartheta }}-1]}^{1-\mathsf{\vartheta }}$
${\mathrm{P}}_0$ and $\mathsf{\vartheta }$	1 × 105 Pa and 0.45
Calculation model for relative permeability in aqueous and gaseous phases ${\mathrm{k}}_{\mathrm{r}\mathrm{A}}, {\mathrm{k}}_{\mathrm{r}\mathrm{G}}$	${\mathrm{k}}_{\mathrm{r}\mathrm{A}}={\left({\displaystyle \frac{{\mathrm{S}}_{\mathrm{A}}-{\mathrm{S}}_{\mathrm{i}\mathrm{r}\mathrm{A}}}{1-{\mathrm{S}}_{\mathrm{i}\mathrm{r}\mathrm{A}}}}\right)}^{\mathrm{n}},$${\mathrm{k}}_{\mathrm{r}\mathrm{G}}={\left({\displaystyle \frac{{\mathrm{S}}_{\mathrm{G}}-{\mathrm{S}}_{\mathrm{i}\mathrm{r}\mathrm{G}}}{1-{\mathrm{S}}_{\mathrm{i}\mathrm{r}\mathrm{G}}}}\right)}^{{\mathrm{n}}_{\mathrm{g}}}$
$\mathrm{n},$${\mathrm{n}}_{\mathrm{g}},$${\mathrm{S}}_{\mathrm{i}\mathrm{r}\mathrm{A}},$ and ${\mathrm{S}}_{\mathrm{i}\mathrm{r}\mathrm{G}}$	3.50, 3.50, 0.30, and 0.03

. Notably, all reservoir parameters listed in Table 1 are obtained from logging data of the SH2 site [46], thereby ensuring that the simulation effectively reflects the actual reservoir conditions. Additionally, the Van Genuchten capillary pressure model [47] and Stone relative permeability model [48] are used to characterize the gas–water flow behavior.

2.4. Simulation Scheme

To systematically evaluate the injection-production performance, a parametric analysis was conducted considering five key factors: fracture half-length (L_F), fracture conductivity (C_F), injection pressure (P_I), injection temperature (T_I), and production temperature (P_P)_. The specific simulation scheme is summarized in

Table 2. Simulation scheme.

Scenario No.	LF, m	CF, D·cm	PI, MPa	TI, °C	PP, MPa	Analysis Objective
1#	—	—	[16, 20]	[30, 90]	[1.5, 7.5]	Reference case
2#	[10, 50]	100	20	60	4.5	The effect of LF (Section 3.1)
3#	[10, 50]	[10, 1000]	20	60	4.5	The effect of CF (Section 3.2)
4#	[10, 50]	100	[16, 20]	60	4.5	The effect of PI (Section 3.4)
5#	[10, 50]	100	20	[30, 90]	4.5	The effect of TI (Section 3.4)
6#	[10, 50]	100	20	60	[1.5, 7.5]	The effect of PP (Section 3.5)

.

2.5. Grid Independence Test

A grid independence test was conducted to determine the grid size, as shown in Figure 2. Notably, the grid density was increased progressively from the top and bottom of the model toward the center to accurately simulate heat and mass transfer within the fracture. The calculation results stabilized at a grid count of 3074, where the grid size in the reservoir ranges between 2 and 5 m. All subsequent simulations were conducted based on this grid scheme to ensure the reliability of the results.

3. Results

Gas production rate (Q_P, defined as the daily CH₄ volume recovered, m³/d) and recovery rate (R_E, defined as the ratio of the cumulative recovered gas mass to the gas mass contained in the hydrate deposits) served as the primary analytical indicator. Additionally, in accordance with China’s offshore development technical specifications, development effectiveness is categorized by R_E, with R_E reached 70% defined as Category I and 85% as Category II [49]. Given this, the average gas production rate (Q_AP, m³/d) at R_E of 0.70 and 0.85 was employed to measure the development efficiency.

3.1. The Effect of Fracture Length

Figure 3 depicts the production performance at different L_F values. For the unfractured scenario (L_F = 0 m), the Q_P value remained below 10 m³/d during the initial 1800 days. This poor recovery efficiency stems from the ultra-low effective permeability of the deposits, which significantly impedes mass and heat transfer between the injection wells and production wells. As a result, this phase constitutes an ineffective thermal stimulation stage, with productivity entirely attributable to depressurization. The reservoir physical field in Figure 4a,b also confirmed this point. Following approximately 1900 days, the thermal stimulation gas was displaced into the production well, with Q_P value gradually increasing to eventually reach 29 m³/d at 3600 days. It is noteworthy that the growth trend of Q_P appears nearly linear on a logarithmic coordinate system, indicating it underwent a relatively stable exponential growth phase. Accordingly, the productivity contribution induced by thermal stimulation gradually became predominant during this phase. Additionally, increasing the L_F value from 10 to 50 m results in a dramatic decrease in the ineffective thermal retardation period (from about 900 to 100 days). This demonstrates that fracturing is crucial for enhancing inter-well connectivity and strengthening inter-well interaction, as it can provide high-conductivity pathways for gas migration from injection wells to production wells. The reservoir physical field in Figure 4c–l can support this point. Furthermore, a distinct productivity depletion phase can be observed following the peak in Q_P. This results from the majority of methane gas having been recovered, leading to the water flooding of production wells. The research conducted by Ma et al. [50] and Zhong et al. [37] also observed this phenomenon in water-driven conditions. In fractured scenarios, the ultimate R_E value ranges from 0.97 to 0.99, whereas in the non-fractured scenario, it is only 0.42. This demonstrates that reservoir reconstruction is crucial for productivity enhancement, even with a limited reconstruction area, which is consistent with the findings of the second pilot production test in the Shenhu Area.

As the L_F value increases from 0 to 50 m, the Q_AP value at R_E of 0.70 increases from 17.80 to 54.55 m³/d. The increase is most pronounced in the 0–10 m range, rising by a factor of 2.16, followed by an approximately linear increase thereafter. Differently, the Q_AP at R_E of 0.85 does not always increase as L_F increases; its value fluctuates between 41.09 and 44.21 when L_F increases from 10 to 40 m. This indicates that the influence of L_F on development effectiveness is not sensitive at this point. The mechanism behind this phenomenon is that under operating conditions where the vast majority of gas has been recovered and water flooding has occurred, L_F has a negligible effect on residue gas recovery. In addition, the Q_AP reaches its maximum value of 49.90 m³/d when L_F is 50 m. This does not imply that fully interconnected wells facilitate residual gas recovery, but rather that timely gas recovery during the initial production phase prevents gas escape. Another interesting phenomenon is that when L_F is 0, 10, and 20 m, the values of Q_AP at R_E of 0.70 are higher compared to those at R_E of 0.85. However, when L_F exceeds 30 m, the opposite is true. This is due to a more pronounced thermal stimulation effect when inter-well connectivity increases, resulting in an earlier peak gas production period. Overall, the recovery efficiency is most ideal when the wells are fully interconnected by fractures.

3.2. The Effect of Fracture Conductivity

Figure 5 depicts the production performance at different C_F values. With increasing C_F from 0 to 250 D·cm, the initial Q_P value (illustrated at production time of 50 days) gradually increases from 8.14 to 70.04 m³/d, and the ineffective thermal stimulation period shortens from 1800 days to 360 days. This indicates that increasing C_F not only enhances productivity by depressurization but also improves thermal stimulation effectiveness. The reservoir gas distribution in Figure 6 further corroborates this observation. However, when L_F increases from 250 to 1000 D·cm, Q_P remains virtually unchanged. This indicates that there exists a critical C_F value for enhancing mass and heat transfer efficiency. The changes in R_E also confirmed this point. For Q_AP at R_E of 0.70, its approximation exhibits exponential growth as C_F increases from 10 to 250 D·cm, after which the growth rate diminishes progressively. Furthermore, the growth rate becomes more pronounced at L_F ≥30 m. Interestingly, when the C_F value exceeds 100 D·cm, the Q_AP at L_F = 50 m is actually lower than that at L_F = 40 m. This indicates that under entirely interconnected well conditions, an excessively high C_F value exacerbates water flooding severity, rendering C_F increases ineffective for enhancing productivity. For Q_AP at R_E of 0.85, its variation trend with C_F is similar. The difference lies in the fact that even at C_F > 100 D·cm and L_F = 50 m, Q_AP remains the most significant. This indicates that full interconnectivity between wells is more conducive to achieving Category III development effectiveness. Overall, scale-up reservoir fracturing is more conducive to enhancing injection-production development efficiency, and it is recommended that the C_F value be no less than 100 D·cm. This fracture scale is highly feasible from an engineering perspective, as existing research has shown that significantly wide fractures can be generated within hydrate deposits [22,25,27,28,51,52].

3.3. The Effect of Injection Pressure

Figure 7 depicts the production performance at different P_I values. Increasing P_I shortens the ineffective thermal stimulation period significantly, specifically, from 800 to 300 days as P_I increases from 16 to 20 MPa. Additionally, the peak Q_AP increased from 73.56 to 108.93 m³/d, and the peak occurrence time reduced from 1210 to 590 days. This indicates that increasing the P_I can enhance production efficiency, as more hot water is injected and the displacement efficiency is improved.

The evolution of R_E reveals that the primary influence zone of P_I is distributed at R_E of 0.35–0.75. The variation in Q_AP further corroborates this argument. It can be observed that Q_AP exhibits near-linear growth with increasing P_I. Particularly when L_F ≥ 20 m, its growth rate at R_E = 0.70 is significantly higher than that at R_E = 0.85. Additionally, we also found that increasing P_I results in premature water flooding when the wells are fully interconnected, which leads to an insignificant improvement in Q_AP at R_E = 0.70. For Q_AP at R_E = 0.85, when L_F exceeds 30 m, its improvement by increasing P_I is also limited. Therefore, in practical applications, attention should be paid to optimizing the P_I value according to production stages. During the initial production phase, P_I can be increased to enhance the water-driven efficiency, while in the later production phase, P_I should be reduced to control water intrusion.

3.4. The Influence of Injection Temperature

Figure 8 depicts the production performance at different T_I values. It can be observed that increasing T_I can enhance development efficiency, as the increased heating efficiency facilitates hydrate decomposition. Specifically, as T_I increases from 30 to 90 °C, the ineffective thermal stimulation period shortens from 707 to 275 days (with the most significant reduction from 30 to 45 °C), the peak value of Q_AP increases from 61.37 to 151.90 m³/d, and the peak time is reduced from 1200 to 500 days. The evolution of R_E indicates that T_I exerts the most pronounced influence during the mid-mining phase (when R_E ranges between 0.30 and 0.70). Additionally, Q_AP exhibits near-linear growth with increasing T_I. Furthermore, as L_F increases, T_I shows a growing tendency to affect Q_AP. An unexpected phenomenon is that when R_E is 0.70, and T_I is lower than 60 °C, the Q_AP at L_F = 50 m is actually lower than that at L_F = 40 m. This is due to premature water flooding, greatly reducing energy utilization efficiency. Additionally, increasing TI can compensate for energy shortages caused by water flooding. As can be seen, when T_I reaches 90 °C, the value of Q_AP is maximized at R_E values of 0.70 and 0.85, corresponding to 109.73 and 98.28 m³/d, respectively, resulting in the most ideal development efficiency.

3.5. The Effect of Production Pressure

Figure 9 depicts the production performance at different P_P values. It can be observed that reducing P_P primarily enhances initial development efficiency (R_E < 0.20). As P_P decreases from 7.5 to 1.5 MPa, the initial Q_P increases from 21.56 to 123.32 m³/d. This enhancement is attributed to the fact that lower pressure facilitates the hydration decomposition around the depressurization well. Furthermore, based on the large-scale experiment conducted by Feng et al. [10], a lower P_P also causes a higher driving force for gas recovery. This results in a shorter ineffective thermal stimulation period. Additionally, Q_AP decreases approximately linearly with increasing P_P. It should be noted that reducing P_I can also induce more severe water flooding. This results in a lower Q_AP at L_F = 50 m compared to that at L_F = 40 m (R_E = 0.70 and P_P ≤ 4.5 MPa). Therefore, it is recommended to adopt a low initial P_P value to enhance development efficiency.

4. Multivariate Importance Evaluation

The above findings reveal that development efficiency is influenced by fracture parameters as well as injection and production parameters. However, the specific weight of each factor in this process has yet to be quantitatively established. Accordingly, this research utilizes eXtreme Gradient Boosting (XGBoost) and SHapley Additive exPlanations (SHAP) algorithm to systematically evaluate the importance of these factors. The XGBoost algorithm, an advanced implementation of gradient-boosted decision trees, enhances predictive performance by aggregating an ensemble of weak learners [53]. In this framework, each new decision tree is trained to compensate for the residuals of the existing ensemble, culminating in a model with high accuracy and robustness. Complementing this predictive strength, SHAP offers a rigorous method for model interpretability, decomposing predictions to reveal the marginal contribution of individual variables [53]. Consequently, the XGBoost-SHAP paradigm has become instrumental in deciphering complex engineering systems, evaluating the relative influence of multiple variables, and providing robust support for decision-making processes [54].

The XGBoost-SHAP framework was implemented through a systematic procedure. First, two separate XGBoost regression models were established, with the Q_AP at R_E of 0.70 and 0.85 designated as the response variables, respectively. The predictor variables consisted of the influencing factors L_F, C_F, P_I, T_I, and P_P. Subsequently, these two models were trained using the simulation dataset, with the training process continuing until the prediction error was reduced to below 0.5%. Finally, the SHAP algorithm was applied to compute the Shapley values for each input parameter, enabling a quantitative assessment of their relative importance. Notably, the dataset consists of 210 groups of simulation results, covering all parameter combinations in the scheme.

The predictive performance of the trained models is illustrated in Figure 10, which presents the prediction errors. It can be found that Q_AP at R_E of 0.70 and 0.85 predicted by the XGBoost models are in close agreement with the simulated values, with prediction errors of 0.45% and 0.35%, respectively. These errors correspond to prediction accuracies exceeding 99.5%, indicating that the dataset size used is sufficient and effective to support reliable prediction and quantitative analysis.

The SHAP values quantifying the contribution of each influencing factor to Q_AP at R_E of 0.70 and 0.85 are presented in Figure 11. The analysis reveals a positive correlation for C_F, L_F, T_I, and P_I, as evidenced by their increasing feature values accompanying higher SHAP values. Conversely, P_P exhibits an inverse relationship. Based on the mean SHAP values, the relative importance of these factors for Q_AP at R_E = 0.70 and 0.85 is ranked in descending order as follows: L_F (15.550 and 9.190), C_F (6.654 and 6.329), P_I (2.908 and 2.174), P_P (2.418 and 2.130), and T_I (2.376 and 1.825). These findings establish L_F as the primary controlling factor for development efficiency, with C_F being secondary. The remaining factors—P_I, P_P, and T_I—display comparable and relatively limited effects.

Figure 12 illustrates the interaction effects among the parameters. The shapviz package in R was used to visualize interaction effects, as shown in Figure 12a,b. To quantify interaction effects, interaction matrices were further plotted as shown in Figure 12c,d. At R_E of 0.70, the interaction strength between L_F and C_F reaches 2.999, which is even greater than the main effect of P_I, P_P, and T_I. This indicates that L_F and C_F exhibit a strong interactive effect on development efficiency and demonstrate positive synergy. Whereas the interaction strengths between other parameters are lower than 1.0, indicating the interaction effect can be ignored. At R_E of 0.85, the interaction strength between L_F and C_F is 1.825, which is higher than the main effect of P_I. As a result, there exists a relatively strong interaction effect between L_F and C_F, though its influence has diminished compared to that at R_E of 0.70.

5. Conclusions

This study evaluated the development efficiency of challenging hydrates under reservoir fracturing and thermal stimulation, and elucidated the synergistic impacts of multivariate using a data-driven approach. The following conclusions can be drawn.

• Reservoir fracturing can significantly improve injection-production behavior as the inter-well mass and heat transfer efficiency is enhanced by the high-conductivity fracture. Specifically, for fracture lengths of 10, 20, 30, 40, and 50 m and fracture conductivity of 100 D·cm, the average gas production at a recovery rate of 0.85 is 2.83, 3.37, 3.30, 3.87, and 4.30 times that of the unfractured scenario, respectively.
• For Class I development standards (with a recovery rate of 0.70), larger-scale fracturing is not always preferable as it may induce more severe water flooding. In our simulations, production performance at a fracture length of 50 m was lower than at 40 m, and fracture conductivity greater than 250 D·cm hardly affects gas recovery.
• Multivariate analysis based on XGBoost-SHAP algorithms revealed that fracture length (SHAP values of 15.550 and 9.190) was the primary factor influencing the development efficiency, followed by fracture conductivity (6.654 and 6.329), injection pressure (SHAP values of 2.908 and 2.174), injection temperature (SHAP values of 2.418 and 2.130), and production pressure (SHAP values of 2.376 and 1.825).
• In plan design, it is essential to promptly adjust the inter-well pressure gradient to control water flooding. Additionally, the synergistic effect between fracture length and fracture conductivity should be considered. The recommended parameters: fracture length = 40 m, fracture conductivity = 100 D·cm.