An integrated method for permeability prediction and fluid identification in tight sandstone reservoirs using geological-prior-guided attention networks: a case study of the X Block, Chang 8 member, Jiyuan oilfield

Abstract

Permeability prediction and sweet-spot identification in tight sandstone reservoirs are challenging because of complex pore–throat structures and strong heterogeneity. This study uses data from 12 cored wells in the Chang 8 Member, Jiyuan Oilfield (western Ordos Basin) to develop a permeability-driven integrated workflow for reservoir evaluation. We first build an SE-ResNet18 model with one-dimensional convolution and residual learning to capture vertical continuity in well logs, achieving R² = 0.86 and RMSE = 0.287 mD for permeability regression. We then design a well-log-based sweet-spot index (Issp) and embed it as a geological prior through an attention-gating mechanism to form a knowledge-guided model (KG-SE-ResNet18). This knowledge guidance improves reservoir-type classification accuracy from 86.95% to 89.28%. Overall, the proposed framework enhances both prediction accuracy and geological consistency, providing a practical approach for fine-scale reservoir evaluation and well-placement optimization in tight sandstones.

Introduction

Permeability is a key parameter controlling petrophysical properties and fluid flow in tight sandstone reservoirs, and thus directly affects production evaluation and enhanced oil recovery^1,2. Core analysis provides reliable measurements, but its sparse sampling, high cost, and limited representativeness restrict its application in highly heterogeneous tight sandstones. In contrast, conventional well logs provide continuous and multi-physics information, making them indispensable for reservoir evaluation. However, permeability prediction methods developed for conventional reservoirs are often less effective in tight sandstones because of complex pore-throat structures and strong heterogeneity^3,4. This issue is particularly evident in the Chang 8 Member of the Jiyuan Oilfield, where permeability varies markedly even at similar porosity, making simple porosity-permeability relationships inadequate^5–9.

Existing permeability prediction methods can generally be grouped into three categories. The first includes physics-based methods, such as FFI and NMR-derived models, which relate permeability to pore-throat characteristics and fluid occurrence mechanisms^10–13. Although physically interpretable, these methods usually require core calibration and often have limited applicability across lithologies and fluid systems¹⁴. The second category includes flow-unit or classification-assisted methods, which partition reservoirs according to pore structure, lithology, grain size, or multivariate clustering to reduce heterogeneity and improve prediction accuracy^15–20. Their effectiveness, however, strongly depends on the reliability of flow-unit classification. The third category includes machine learning and deep learning methods. Tree-based ensemble models such as XGBoost and GBDT can effectively capture nonlinear relationships between permeability and well-log responses^21–27, and their performance can be further improved by hyperparameter optimization methods such as IPSO^28–30. However, most of these models treat samples independently and do not explicitly capture vertical continuity in well logs. In contrast, one-dimensional convolutional networks are better suited to learning depth-dependent patterns, and recent studies have shown their potential in predicting tight-reservoir properties from log data^31,32.

Despite these advances, two issues remain. First, permeability prediction and fluid identification are usually treated as separate tasks, which weakens the transfer of pore-structure information into fluid-type recognition. Second, purely data-driven classifiers may produce geologically inconsistent results in complex intervals because geological constraints are not explicitly incorporated. To address these issues, this study proposes an integrated physics-data-driven framework that links petrophysical characterization and fluid identification through a cascaded learning strategy. Specifically, an SE-ResNet18 regression model is first developed to extract depth-dependent features from one-dimensional well logs for continuous permeability prediction^33–35. The predicted permeability is then introduced into the subsequent classification stage as an endogenous physical constraint. In addition, a well-log-based sweet spot index (Issp) is designed as prior geological knowledge and embedded through a dynamically weighted attention-gating mechanism, forming the KG-SE-ResNet18 model and improving geological consistency in complex intervals^36,37.

The main contributions of this study are as follows. First, we develop an SE-ResNet18-based sequential permeability prediction model that captures the vertical continuity of well logs through one-dimensional convolution and channel attention. Second, we construct a permeability-driven cascaded workflow in which predicted permeability is used as an endogenous physical constraint for fluid identification. Third, we introduce the geological prior index Issp into KG-SE-ResNet18 through an attention-gating mechanism, enabling knowledge-guided feature learning and improving geological consistency. Overall, the proposed framework provides an integrated workflow for permeability prediction and fluid identification in tight sandstone reservoirs.

Geological settings

The Jiyuan Oilfield is located in the central–western Ordos Basin, spanning the Tianhuan Depression and the northern Shaanxi Slope. The overall structure is a gently southwest-dipping monocline with local nose uplifts and minor anticlines (Fig. 1a)³⁸. This relatively stable structural setting provides a consistent regional framework, while local uplifts may introduce subtle heterogeneity relevant to permeability and fluid distribution. The target interval is the Chang 8 Member of the Triassic Yanchang Formation, buried at depths of 2000–3000 m, characterized by low porosity and ultra-low permeability tight sandstones. Deposited during the middle Late Triassic, Chang 8 belongs to a shallow-water delta system dominated by delta-front subfacies. Reservoirs mainly occur in subaqueous distributary channels and mouth bars, with sandbody distribution controlled by paleogeomorphology and hydrodynamics. For this study, the key implication is that channel and mouth-bar microfacies form laterally continuous but vertically stacked sand bodies with distinct reservoir qualities, which directly motivates sequence-based modeling of well logs. Such microfacies heterogeneity intensifies reservoir complexity and complicates permeability prediction.

Fig. 1.

(a) Structural location map of the Ordos Basin. (b) Cross-plot of core permeability versus porosity in the study area.

The reservoir is dominated by feldspathic litharenite and lithic sandstone. To focus on modeling-relevant factors, we summarize the petrophysical controls as follows: (i) a complex pore system dominated by fine pores and micro-throats, and (ii) strong heterogeneity caused by microfacies variation, diagenesis, and local microfractures^39–44. Core analyses reveal porosity ranging from 2% to 13% (average ~ 10.4%) and permeability mainly between 0.01 and 5 mD (average ~ 0.53 mD), consistent with typical low-porosity, ultra-low-permeability reservoirs. These geological factors also imprint on log responses: higher shale/mud content in mouth-bar or argillaceous intervals typically increases GR and may reduce RT, whereas tighter/compacted sandstones tend to show higher DEN and lower AC; dissolution and microfractures can locally enhance flow capacity but may produce subtle or mixed log signatures^39,42,43. Therefore, permeability and fluid-type discrimination requires integrating multiple logs and their vertical context, rather than relying on a single log or a simple porosity–permeability trend.

Figure 1(b) illustrates the core porosity–permeability cross-plot for the study area and highlights the weak and highly scattered relationship between the two variables. Under identical porosity, permeability may vary by one to two orders of magnitude (Fig. 1b), indicating that, besides porosity, micro-pore structure, lithology, and diagenesis are critical controls. This pronounced scatter directly motivates the use of multi-log nonlinear learning for permeability prediction and the incorporation of geological prior knowledge to reduce classification ambiguity in complex intervals.

Research data and methods

Data source and processing

The experimental data for this study were sourced from conventional logging and core analysis results of 12 core wells in the A block. A dataset was created using data from 11 wells, consisting of 2308 sample points (Table 1), with the remaining well used for independent validation. Each sample includes six types of conventional logging features: acoustic time difference (AC), compensated neutron logging (CNL), density logging (DEN), natural gamma logging (GR), spontaneous potential logging (SP), and resistivity logging (RT).

Table 1.

Partial training sample data.

AC(µs·m⁻¹)	RT(Ω·m)	DEN(g·cm⁻³)	GR(API)	SP (mV)	CNL (%)	PERM (mD)
222.81	62.595	2.525	82.56	89.211	21.976	0.421
220.20	71.662	2.56	83.482	89.057	21.225	0.237
214.41	128.984	2.532	72.474	91.46	14.862	0.314
217.11	128.421	2.507	74.927	91.95	16.014	0.482
218.29	208.857	2.491	67.007	94.141	18.435	0.615
217.36	139.786	2.505	64.088	94.706	17.494	0.49
217.71	93.206	2.511	65.167	95.255	16.803	0.464
220.63	71.349	2.504	72.057	95.768	16.731	0.571

The dataset used in this study was collected from 12 cored wells in the A block. Of these, 11 wells were used for model development, yielding 2308 samples, whereas the remaining well (G16) was fully held out as an independent blind well for external validation. In the main experiments, both regression and classification were performed on pooled samples from the 11 development wells. Repeated 5-fold stratified cross-validation was used for the regression stage, while a sample-level train-test split was adopted for the classification stage. All logging features were normalized using parameters fitted only on the training set, which were then applied to the test set. Samples with missing target permeability values were removed, while missing values in the logging curves, if any, were imputed using the median of the corresponding log calculated from the training set. All samples were collected from a depth interval of 2500–2800 m. Approximately 62.8% of the permeability values are lower than 1 mD and 44.0% are lower than 0.5 mD, whereas only about 7.9% exceed 5 mD, indicating a typical tight/ultra-low-permeability reservoir dominated by low-permeability samples. The well log data at each depth point were treated as a 1D sequence of length N, forming an input tensor of shape (1, N) to fully capture the potential relationships along depth (vertical) and across logging channels. Here, the sequence length N represents the vertical context window (i.e., the effective receptive field) used by the 1D-CNN to learn depth-continuity patterns. A smaller N tends to preserve high-frequency variations and is more responsive to thin interbedded layers, but may be more sensitive to noise; a larger N strengthens continuity learning and stabilizes predictions, while potentially smoothing very thin beds whose thickness is shorter than the effective receptive field. In practice, N can be selected by relating the expected minimum resolvable bed thickness to the log sampling interval (converting thickness to the number of samples) and further tuned through sensitivity tests to balance thin-bed resolution and prediction stability. This approach preserves the independent contribution of each curve while allowing the convolutional model to capture their local interactions, thereby enhancing the model’s ability to express geological nonlinearities.

Permeability prediction model based on SE-ResNet18

Well logs are ordered depth series. One-dimensional convolution is suitable because it learns local patterns along depth and aggregates them into higher-level features, thereby capturing vertical continuity that is often overlooked when treating samples independently. Traditional deep networks may become difficult to train when they are very deep. ResNet alleviates this issue by adding shortcut connections, which allow information to flow more easily through the network and improve feature learning. Based on ResNet-18, we introduce the Squeeze-and-Excitation (SE) block to adaptively emphasize more informative logging channels for permeability prediction. In practice, the SE block can be understood as a learnable “channel weighting” module: it assigns a higher weight to logs that contribute more to prediction and suppresses less relevant ones.

1

2

3

The SE block operates in three intuitive steps (Eqs. (1)–(3)). First, it summarizes each logging channel into a single representative value. Second, it converts these summaries into channel weights between 0 and 1 through a simple gating function. Third, it rescales the original features using these weights. As a result, informative channels are amplified and less relevant channels are suppressed, which helps the network focus on permeability-sensitive logging responses.

Given the limited sample size of the well logging dataset, to prevent overfitting due to overly complex network structures while ensuring both model performance and training stability, ResNet-18, which is structurally simple yet empirically effective, was selected as the backbone network (Fig. 2). Under small-sample conditions, ResNet-34 exhibits noticeable overfitting, whereas ResNet-18 demonstrates stable performance, validating its suitability for the current dataset size. The selection of ResNet18 reflects a deliberate trade-off between model capacity and data sparsity, ensuring stable convergence and avoiding overfitting under limited core-calibrated datasets.

Fig. 2.

Comparison of training process loss curves between ResNet-18 and ResNet-34. (a) loss curve of ResNet-18. (b) loss curve of ResNet-34.

On this basis, the SE-ResNet18 model was constructed (Fig. 3). This model uses ResNet-18 as the backbone network, and in each basic residual block, an SE module is inserted after the second convolutional layer and before the addition of the identity mapping, allowing dynamic adjustment of channel weights.

Fig. 3.

SE-ResNet18 network architecture diagram.

The well log sequence data are first fed into a one-dimensional convolution layer and a max-pooling layer to perform initial feature extraction and downsampling. The features then pass through four residual stages (Block1–Block4), each consisting of two residual blocks embedded with SE modules. The feature maps subsequently undergo global average pooling, a Dropout layer, and a fully connected layer, finally producing the permeability prediction output. This architecture emphasizes critical channel information throughout multi-level feature extraction, thereby enhancing prediction accuracy. The model is optimized using the Adam optimizer, combined with a cosine annealing schedule for dynamically adjusting the learning rate⁴⁵; a higher learning rate is maintained at the early stage to accelerate convergence, and gradually decreased in later stages to improve training stability.

The main training settings and regularization strategies of the SE-ResNet18 model are summarized below (Table 2). In addition, 20% of the training data were used as a validation subset for model selection and learning-rate adjustment. To obtain a more reliable estimate of model performance under limited data conditions, repeated 5-fold stratified cross-validation was further conducted, and the results are reported as the mean ± standard deviation of R² and RMSE.

Table 2.

Comparison of performance of different models on the test set.

Parameter	Value
Optimizer	Adam
Initial learning rate	7 × 10⁻⁴
Weight decay	7 × 10⁻⁴
Batch size	32
Maximum epochs	40
Loss function	Smooth L1 loss (Huber loss)
Dropout rate	0.3
Early stopping	Validation loss, patience = 10, δ = 5 × 10⁻⁵
Validation protocol	Repeated 5-fold stratified cross-validation

Construction of prior knowledge

To quantitatively characterize the ability to identify reservoir “sweet spots,” this study constructs a well logging sweet spot comprehensive index based on geological response mechanisms and heuristic factors. It should be noted that Issp is an interpretable, well-log-derived index designed to summarize several key petrophysical controls related to reservoir quality. Rather than claiming to fully represent all geological constraints, Issp provides a practical proxy that reflects the combined effects of storage capacity, fluid mobility, and shale content on potential productivity. In this study, Issp is used as a soft prior feature to guide the subsequent learning process. The index integrates permeability (PERM), porosity (POR), true resistivity (RT), natural gamma (GR), acoustic transit time (AC), and water saturation (SW), calculated as follows:

4

where α is a scaling factor used to balance the permeability term with other logs in the index, and ϵ is a small positive number to avoid division by zero. In this study, α was determined empirically on the training wells by testing several candidate values and selecting the one that yields the clearest separation of Issp distributions among the four interpreted reservoir types (oil, oil–water, water, and dry), while keeping Issp values within a comparable numerical range. The logs included in Issp (PERM, POR, RT, GR, AC, and SW) were selected because they represent complementary physical controls: PERM and POR characterize storage/flow capacity, RT and SW reflect fluid-related electrical responses, and GR and AC are sensitive to shale content and compaction/rock framework. Parameters favorable for reservoir quality (PERM, POR, RT) are placed in the numerator, whereas parameters detrimental to reservoir quality (GR, AC, SW) are placed in the denominator. In Issp, permeability is expressed in logarithmic form to compress its dynamic range, whereas water saturation is introduced as an exponential term in the denominator to enhance its penalizing effect on reservoir quality. The scaling factor α was determined empirically using only the training wells, while ϵ was defined as a small positive constant for numerical stability. To validate the discriminative capability of Issp, a statistical analysis was conducted on the main stratigraphic intervals based on well log interpretation and production data. Reservoirs were classified into four categories according to fluid type: oil zone (sweet spots), oil-water zone (oil-water transition zone, water zone with oil shows), water zone, and dry zone.

The Issp values vary significantly across different reservoir types (Table 3). Oil zone exhibit the highest indices, followed by oil-water zone, with water zone being lower and dry zone the lowest. This indicates that the latter possess poorer petrophysical and electrical properties, representing ineffective reservoirs. The Issp ranges and mean values for each reservoir type are as follows: oil zone, 8.5–62.3 (mean 28.7); oil-water zone, 5.2–23.5 (mean 10.2); water zone, 3.5–6.1 (mean 4.8); and dry zone, 0.3–3.8 (mean 2.3). Based on the observed Issp statistics in the training wells (ranges and mean values) and the interpreted fluid-type labels, we set preliminary cutoffs at the overlap-minimizing boundaries between adjacent classes. To avoid label leakage, the Issp thresholds were determined solely from the training wells and kept fixed for validation, test, and blind-well analyses. Specifically, Issp ≥ 18 indicates oil zone; 6.0–18 corresponds to oil–water zone; 3.5–6.0 corresponds to water zone; and Issp < 3.5 corresponds to dry zone. These thresholds are intended as practical criteria for this block and may require recalibration when applied to other reservoirs with different depositional settings or fluid systems. Moderate perturbations in α or in the threshold values mainly affect samples near class boundaries, particularly within oil–water transitional intervals, whereas the separation between end-member classes remains relatively stable. This indicates that the sensitivity of Issp is mainly concentrated in transitional zones. For applications to other blocks/reservoirs, the Issp thresholds should be recalibrated using a small set of core-/test-calibrated intervals or local production-based interpretations to account for possible domain shifts in facies, diagenesis, and fluid properties. Overall, these cutoffs provide a useful reference for rapid sweet-spot screening in the studied block.

Table 3.

Comprehensive logging-based sweet spot indexes for different reservoirs.

AC (µs·m⁻¹)	RT (Ω·m)	GR (API)	PERM (mD)	POR (%)	SW (%)	Interpreted result	Issp
219.36	192.62	69.23	1.08	11.44	22.38	oil zone	36.95
233.22	113.09	73.90	1.44	14.06	23.67	oil zone	28.14
229.74	114.40	81.26	1.15	13.36	24.26	oil zone	21.31
239.29	49.50	67.99	2.64	13.47	42.21	oil-water zone	15.07
244.48	28.39	71.85	4.27	14.57	51.54	oil-water zone	10.15
233.08	16.11	68.87	4.48	14.23	67.07	water zone	5.39
240.61	11.93	69.28	6.96	15.73	73.16	water zone	4.88
226.26	90.47	81.38	0.29	8.44	49.83	dry zone	2.76
216.45	71.74	76.12	0.13	6.60	48.98	dry zone	0.92

Knowledge-guided SE-ResNet18 model (KG-SE-ResNet18)

Model concept

In tight sandstone intervals with subtle log contrasts, using Issp alone may be insufficient for reliable reservoir-type identification. Although Issp integrates multiple logging parameters, including petrophysical and electrical properties, and demonstrates certain static discriminative capability, its empirically driven nature limits its ability to capture nonlinear couplings and multi-scale response features. This limitation becomes particularly evident when the oil–water contact is indistinct or when the differences between low-efficiency and effective oil layers are subtle, leading to a significant decline in classification accuracy.

To address these issues, this study proposes an integrated cascaded framework centering on the KG-SE-ResNet18 model. Unlike conventional parallel modeling, this approach explicitly links petrophysical inversion with reservoir classification by adopting a sequential hierarchical architecture (Fig. 4). The framework consists of two coupled stages: the Feature Extraction & Regression Stage (SE-ResNet18) for continuous permeability prediction, and the Knowledge-Guided Classification Stage (KG-SE-ResNet18) for fluid identification.

Fig. 4.

Flowchart of the knowledge-guided method for reservoir classification.

Implementation steps

The deep feature extraction module employs the pre-trained SE-ResNet18 as the backbone network, encoding the input logging sequence x_main to extract deep features f_deep. In the classification stage, the input feature space is significantly expanded beyond raw logging curves. The predicted permeability (PERM_pred) obtained from the preceding regression stage is integrated as an ‘endogenous physical parameter’ into the feature mapping process. This cascading design ensures that the classification model inherently ‘perceives’ the reservoir’s pore-structure characteristics, effectively simulating the geological expert’s heuristic of ‘evaluating storage capacity before determining productivity’.

The knowledge-attention fusion module is the core innovation of the model. By introducing the prior Issp index, it dynamically fuses deep features with prior knowledge. The KG-SE-ResNet18 model does not simply concatenate raw data; it employs an attention-gate mechanism where the Issp index and PERM_pred act as dual prior constraints. The deep features f_deep, now embedded with physical information from PERM_pred, are fused with x_issp to form the input. This coupling mechanism allows the network to adaptively reweight deep sequential features based on petrophysical logic, ensuring that the fluid identification results remain geologically consistent. This fusion module learns a gating coefficient α∈(0,1) to control how much the model should rely on the prior index under different log responses. When the log signatures are clear, the model primarily uses the deep features; when the interval is ambiguous, the gate can increase the contribution of Issp to improve geological consistency. In other words, prior knowledge is used to guide feature learning rather than being simply appended as an extra input⁴⁶.

5

where α∈(0,1) is a learnable gating weight that controls how strongly the prior knowledge contributes to classification. The details of the gating function are given in Eq. (5), but conceptually it produces a context-dependent weight from the deep features. The weighted prior feature is then obtained by scaling x_issp with α, as shown in Eq. (6).

6

Finally, the model merges two sources of information: the deep sequential features f_deep learned from logs and the weighted prior feature α ⋅ x_issp ​. This fused representation is then fed into a lightweight classifier to output the reservoir-type label. This fused vector is then fed into a multilayer perceptron classifier composed of fully connected layers, batch normalization (BN), ReLU activation, and Dropout, which outputs the final reservoir type prediction. The adaptive coefficient α enables the model to dynamically balance data-driven features and geological prior knowledge. In relatively homogeneous intervals, the model relies more on deep sequential features, whereas in complex oil–water transition zones or ambiguous logging responses, the contribution of Issp is automatically enhanced, mimicking expert-driven geological reasoning. In a field application, engineers can run the model in two passes: (i) input the required log suite (AC, RT, DEN, GR, SP, CNL, and any additional logs used for Issp) to obtain continuous PERM_pred; (ii) compute Issp and feed (PERM_pred, Issp, f_deep) into the classifier to output reservoir types along depth. The outputs can be used to quickly screen sweet spots, delineate oil–water transition intervals, and support perforation and well-placement decisions. When core permeability is unavailable, PERM_pred can directly replace PERM in the Issp calculation for a consistent workflow.

Experimental results and analysis

Permeability prediction results

Model performance comparison and analysis

To evaluate the effectiveness of the proposed method, five regression approaches were compared: ResNet18, SE-ResNet18, BP neural network, XGBoost, and conventional regression (Ridge regression)^47–49. All models were trained on the same dataset under identical conditions. To improve the reliability of performance evaluation under limited data conditions, repeated 5-fold stratified cross-validation was conducted. This cross-validation was performed on pooled samples from the 11 development wells and therefore reflects sample-level rather than well-level model comparison. For each model, the validation metrics obtained from the five folds were averaged within each repetition, and the final results were reported as the mean ± standard deviation across repeated runs. The comparative performance of the five models in terms of R² and RMSE is presented below (Table 4). Among the tested models, SE-ResNet18 achieved the highest R², the lowest RMSE, and the smallest standard deviation, indicating both superior predictive accuracy and stronger stability.

Table 4.

Performance comparison of the five regression models for permeability prediction.

Model	R 2	RMSE(mD)
SE-ResNet18	0.85–0.87(0.86, 0.007)	0.274–0.316(0.287, 0.019)
ResNet18	0.82–0.84(0.83, 0.009)	0.337–0.391(0.363, 0.021)
BP	0.78–0.80(0.79, 0.011)	0.433–0.467(0.449, 0.013)
XGBoost	0.79–0.82(0.81, 0.013)	0.386–0.458(0.417, 0.026)
Conventional Regression	0.67–0.70(0.69, 0.035)	0.657–0.794(0.717, 0.054)

0.85–0.87 (0.86, 0.007) denotes the minimum-maximun range (mean value, standard deviation) obtained from repeated 5-fold stratified cross-validation.

The scatter plots of predicted versus measured permeability (Fig. 5) confirm this advantage: SE-ResNet18 shows the closest agreement with the 1:1 line and achieves a 98% conformity rate within half an order of magnitude, outperforming baseline regression and machine-learning methods.

Fig. 5.

Cross-plot of Predicted vs. Measured Permeability in the Study Area. (a) Comparison of Results between Conventional Regression and SE-ResNet18. (b) Comparison of Results between ResNet18 and SE-ResNet18. (c) Comparison of Results between BP and SE-ResNet18. (d) Comparison of Results between XGBoost and SE-ResNet18.

Residual analysis in log-permeability space shows that the errors are generally centered around zero, but are not strictly homoscedastic across the full permeability range (Fig. 6). A slight overestimation is observed at the low-PERM end, whereas a mild underestimation appears at the high-PERM end, with somewhat larger dispersion in the high-permeability interval.

Fig. 6.

Residuals of permeability prediction in log-permeability space.

Model interpretability analysis

In this study, the importance of input features was quantitatively assessed from two perspectives. First, during model training, the SE module treated the six well logging curves (AC, RT, DEN, GR, SP, and CNL) as independent channels. By learning attention scores and normalizing them into weights, the module directly reflected the model’s reliance on different features. As shown in Fig. 7a, acoustic transit time (AC) received the highest attention weight (0.24). Second, after training, permutation feature importance was applied for validation. This method evaluates feature contribution by randomly shuffling each feature in the validation set and measuring the corresponding reduction in R². The larger the decline in R², the more important the feature. The analysis (Fig. 7b) indicates that shuffling AC led to the greatest drop in R², with a normalized importance of 0.33.

Fig. 7.

(a) Radar chart comparison of normalized attention weights and permutation importance (b) Permutation feature importance analysis.

Both approaches yielded consistent feature rankings. AC consistently ranked first, followed by density (DEN). Differences emerged between RT and SP: in the attention mechanism, SP ranked higher than RT, whereas in permutation analysis, RT showed greater importance. This suggests that the two features play distinct roles during training and prediction. This discrepancy is expected because attention weights and permutation importance quantify feature relevance from different perspectives. The SE attention weights reflect how the network allocates channel emphasis during representation learning and may be affected by feature redundancy/correlation and interactions among logs. In contrast, permutation importance measures the marginal impact of each feature on predictive performance after training (i.e., the loss of information when a feature is randomized), and therefore tends to highlight features with higher unique contribution under correlated inputs. GR and CNL consistently ranked last in both analyses, indicating limited contribution to permeability prediction.

Validation on an unmodeled well

The trained SE-ResNet18 model was applied to predict the permeability of the G16 well, which was not involved in model training (Fig. 8). The results indicate that the model significantly outperforms conventional regression methods in key reservoir intervals. In high-permeability zones of the main producing layers and low-permeability zones of interbedded layers, SE-ResNet18 predictions closely match the measured permeability curves, accurately capturing abrupt changes and subtle variations in thin sand layers.

Fig. 8.

Permeability prediction results for Well G16 in the study area.

Error analysis of 158 sample points in the G16 well shows that 97.5% of predictions by SE-ResNet18 fall within ± 1 mD, exceeding the 60.4% achieved by traditional regression methods by 37.1%. Additionally, only 4 points exhibit prediction errors larger than ± 1 mD, compared with 63 points for conventional regression methods. These results demonstrate that SE-ResNet18 provides high accuracy and strong stability in permeability prediction, fully meeting the engineering requirements of fine-scale reservoir characterization in field development and effectively addressing the limitations of conventional methods in complex reservoirs.

Reservoir classification results

Model performance comparison and analysis

Based on the original dataset, six input features were used for fluid-type characterization: acoustic transit time (AC), resistivity (RT), natural gamma (GR), porosity (POR), permeability (PERM), and water saturation (SW). In addition, the well logging sweet spot comprehensive index (Issp) was incorporated as geological prior knowledge. Five experimental groups were considered: the proposed KG-SE-ResNet18; the baseline SE-ResNet18; an ablation baseline, SE-ResNet18 (Issp-concat), in which Issp was directly concatenated to the classifier input without the attention-gating mechanism; and two traditional machine learning models (XGBoost and SVM⁵⁰ that also incorporate geological prior knowledge. The five reservoir classes are moderately imbalanced in the dataset. Specifically, the numbers of samples are 507 for oil, 461 for oil–water transition, 470 for water zone with oil shows, 396 for water, and 474 for dry. Although the imbalance is not severe, the water class contains relatively fewer samples than the others. The main classification comparison reported here was based on a sample-level train-test partition of the pooled samples from the 11 development wells,

The test results of the trained SE-ResNet18 and KG-SE-ResNet18 models were compared using confusion matrices (Fig. 9). After introducing knowledge guidance, the overall classification accuracy increased from 86.95% to 89.28%, with the number of correctly identified samples rising from 401 to 412. The recall rates for oil zone, oil–water transition zone, and water zone with oil shows increased from 90%, 81%, and 68% to 93%, 82%, and 76%, respectively, with the identification performance of water zone with oil shows improving most significantly by 8%. This class benefits most from knowledge guidance because its log responses are often intermediate and easily confused with oil–water transition or low-efficiency oil layers when using purely data-driven features. By injecting Issp and permeability-related constraints, the model receives an explicit cue about storage capacity and fluid mobility, which helps reduce ambiguity near class boundaries and improves recall for this transitional class.

Fig. 9.

(a) Classification results of the baseline model, SE-ResNet18. (b) Classification results of the model, KG-SE-ResNet18.

The comparative results (Table 5) show that KG-SE-ResNet18 achieves the highest overall accuracy (89.28%) and F1 score (82.73%). Although SE-ResNet18 (Issp-concat) improves upon the baseline SE-ResNet18, KG-SE-ResNet18 still yields superior performance, indicating that the observed gain arises not only from the inclusion of Issp, but also from the proposed attention-gating mechanism for knowledge fusion. Compared with XGBoost and SVM, which also incorporate prior knowledge, KG-SE-ResNet18 exhibits a more pronounced advantage in complex reservoir classification.

Table 5.

Comparison of performance of different models on the test set.

Model	Accuracy (%)	F1Score (%)	Correctly classified samples
KG-SE-ResNet18	89.28	82.73	412
KG-SE-ResNet18 (PERMpred)	88.12	80.65	407
SE-ResNet18 (Issp-concat)	87.67	81.21	405
SE-ResNet18	86.95	78.97	401
XGBoost	83.94	75.95	387
SVM	81.19	72.29	375

To provide a more diagnostic evaluation under potential class imbalance, this study further supplements class-wise classification metrics, including Precision and Recall, for each reservoir type (Table 6). To further quantify the impact of replacing core permeability with predicted permeability in the classification stage, KG-SE-ResNet18 (PERM_pred) was additionally evaluated on the test set of all development wells, with the results reported (Tables 5 and 6). Compared with the model using core permeability, only slight decreases in Accuracy and F1 were observed, while class-wise Recall remained generally stable, with moderate reductions mainly in transitional classes. No obvious systematic bias toward the oil-zone class was found, indicating that predicted permeability can effectively substitute for core permeability when core data are unavailable.

Table 6.

Comparison of performance of different models on the test set.

Model	KG-SE-ResNet18		KG-SE-ResNet18 (PERMpred)		SE-ResNet18		XGBoost		SVM
	P.	R.	P.	R.	P.	R.	P.	R.	P.	R.
Oil	0.94	0.93	0.93	0.92	0.91	0.90	0.84	0.85	0.76	0.78
OWT	0.77	0.82	0.75	0.79	0.75	0.77	0.71	0.77	0.66	0.71
WOS	0.82	0.76	0.79	0.73	0.75	0.68	0.67	0.62	0.61	0.54
Water	0.69	0.78	0.68	0.77	0.72	0.81	0.81	0.75	0.72	0.66
Dry	0.89	0.88	0.88	0.87	0.90	0.89	0.83	0.85	0.74	0.77

P. denotes Precision; R. denotes Recall; OWT denotes oil–water transition zone; WOS denotes water zone with oil shows.

For reservoir classification, the present study focuses on discrete performance metrics, including Accuracy, F1 score, Precision, and Recall, whereas probability calibration was not explicitly evaluated. In practical deployment, uncertainty-aware strategies such as Monte-Carlo Dropout, deep ensembles, and post hoc calibration may be introduced to quantify prediction confidence, especially in transitional intervals with ambiguous log responses.

Single-well application example

To evaluate the applicability of the model in practical reservoir identification, the blind-well classification prediction was performed on Well G16, which was not included in the training set. The predicted results were compared layer by layer with the integrated geological interpretation derived from well logs, drilling records, and test production data (Fig. 10).

Fig. 10.

Prediction results of reservoir types in Well G16 of the study area. KG-SE-RN18 (KG-SE-ResNet18) and SE-RN18 (SE-ResNet18); “KG-SE-RN18 (SE-RN18 PERM)” indicates the reservoir type prediction from the KG-SE-RN18 model using the SE-ResNet18-predicted permeability as an input feature.

In the main oil-bearing interval from 2665.4 to 2673.7 m, geological interpretation indicates a continuous oil layer. KG-SE-ResNet18 accurately captured the connectivity of this interval, with classification results consistent with the geological interpretation. In contrast, SE-ResNet18 divided it into scattered oil layers and dry layers, disrupting the continuity. In the oil–water transitional zone from 2648.3 to 2654.4 m, SE-ResNet18 predicted a mixture of oil, dry, and oil–water layers with unclear boundaries, whereas KG-SE-ResNet18 stably identified it as an oil–water layer. This demonstrates that the knowledge-guided mechanism enhances the model’s adherence to geological rules, producing results that are geologically reasonable, continuous, and more aligned with actual features, while maintaining stable boundary discrimination and classification consistency.

A brief quantitative summary for blind Well G16 is presented (Table 7). KG-SE-ResNet18 achieves the best performance, with 88.53% Accuracy, 81.22% F1, 0.90 Oil Recall, and 0.76 OWT Recall. Only slight degradation is observed when predicted permeability is used instead of core permeability, confirming the robustness of the proposed workflow. Furthermore, when the SE-ResNet18 predicted permeability was used as input instead of core-measured permeability, KG-SE-ResNet18 still successfully identified the main oil interval (2665.4–2673.7 m), consistent with geological interpretation. This experiment validates the effectiveness of predicted permeability and indicates that integrating predicted permeability can improve reservoir classification reliability when core data are unavailable.

Table 7.

Quantitative summary on the blind well (agreement with interpreted reservoir types).

Model	Accuracy (%)	F1Score (%)	Oil Recall	OWT Recall
KG-SE-ResNet18	88.53	81.22	0.90	0.76
KG-SE-ResNet18 (SE-ResNet18 PERM)	87.35	78.65	0.88	0.73
SE-ResNet18	86.95	77.67	0.87	0.71

Discussion

This study demonstrates that both SE-ResNet18 and KG-SE-ResNet18 achieve substantial improvements in permeability prediction and reservoir classification in tight sandstone reservoirs. The performance advantages primarily stem from the ability of one-dimensional convolutional residual networks to capture depth-dependent sequential patterns in well logging data, which reflect the vertical continuity of sedimentary processes and diagenetic variations. In addition, the SE attention module adaptively reweights informative channels, enabling the model to focus on physically meaningful logging responses.

Compared with conventional regression approaches and traditional machine learning models such as XGBoost and BP neural networks, the proposed models exhibit superior prediction accuracy and stability, particularly in heterogeneous intervals. Feature contribution analysis indicates that acoustic transit time and density are the dominant controlling factors for permeability prediction. This observation is consistent with their strong geological sensitivity to pore structure evolution and compaction intensity, thereby confirming the physical interpretability and geological plausibility of the model outputs. Compared with sample-independent ML workflows, the explicit use of vertical context in SE-ResNet18 provides a more consistent characterization of depth-evolving log responses, which helps reduce local prediction volatility in thinly interbedded or rapidly varying intervals.

By incorporating the comprehensive sweet spot index (Issp) as prior geological knowledge, the KG-SE-ResNet18 model further enhances reservoir classification performance. The ablation results further indicate that the performance gain is not merely due to the inclusion of Issp, but also to the proposed attention-gating design for more effective knowledge-guided feature fusion. This design is particularly effective in complex intervals, such as oil–water commingled and oil-bearing water layers, where conventional data-driven models often suffer from ambiguous log responses and reduced classification reliability. Moreover, the permeability predicted by the SE-ResNet18 model serves as a physically meaningful constraint for the subsequent classification task, ensuring that fluid identification is grounded in pore-structure characterization and forming a permeability-driven integrated workflow.

Practical deployment challenges should be considered before field application. In particular, the workflow requires consistent preprocessing across wells (depth matching, environmental correction, and normalization), because changes in logging tools, sampling intervals, or curve-editing practices can introduce distribution shifts and reduce performance. A practical implementation is to standardize preprocessing and define a minimum required log suite for inference, while allowing automated handling of missing curves when applicable.

The framework is also sensitive to data quality. Noise spikes and cycle skipping in AC, borehole-condition effects on DEN/CNL, shoulder-bed effects on resistivity in thin interbeds, and log–core depth misalignment may lead to local discontinuities in predicted permeability and class confusion near transition zones. Therefore, routine quality control (outlier detection, robustness filtering for isolated spikes, and flags for poor borehole conditions or incomplete logs) is recommended to improve reliability.

Despite these advantages, the generalization capability of the proposed models may be influenced by the limited size of the available dataset and the empirical nature of Issp construction⁵¹. In particular, the dataset used in this study is collected from a single block; therefore, the reported performance mainly reflects within-block (inter-well) generalization. Although the KG-SE-ResNet18 architecture may be transferable to reservoirs with similar logging responses and geological settings, the Issp formulation and its thresholds remain formation-specific unless validated elsewhere. For other tight sandstones, local recalibration may be sufficient, whereas for carbonates or different salinity regimes, more substantial reformulation may be required.

The threshold values of Issp may require recalibration when applied to reservoirs with different depositional settings or fluid systems. When extending the model to other blocks or reservoirs, its transferability is expected to be higher under comparable depositional settings, diagenetic evolution, fluid systems, and consistent logging suites; otherwise, distribution shifts (e.g., facies changes or different pore-throat structures) may lead to degraded performance, necessitating recalibration and/or fine-tuning with a small set of core-calibrated samples. Future work will focus on incorporating multi-scale logging and core data, as well as adopting transfer learning and domain adaptation strategies, to further improve model robustness and enhance its applicability in broader engineering scenarios. To make future work more actionable, we will (i) perform cross-block validation using a small number of cored wells and apply transfer learning/domain adaptation to quantify and reduce domain shift; (ii) report class-wise precision and recall to diagnose class imbalance and error modes in complex oil–water intervals; and (iii) add uncertainty estimation (e.g., ensemble-based confidence intervals) to support risk-aware interpretation and engineering decision-making.

Conclusions

(1) A permeability prediction model based on SE-ResNet18 is developed by integrating one-dimensional convolutional residual learning with channel-wise attention. The model captures the vertical continuity of well logs and provides accurate, stable permeability prediction in tight sandstone reservoirs, outperforming conventional regression and common ML baselines.

(2) Feature contribution analysis reveals that acoustic transit time and density are the dominant well logging parameters controlling permeability prediction, which is consistent with the geological mechanisms governing pore structure development and compaction, while GR and CNL exhibit relatively weaker influence.

(3) By incorporating the well logging sweet spot comprehensive index (Issp) as prior geological knowledge, the KG-SE-ResNet18 model realizes dynamic fusion of deep features and geological constraints through an attention-gating mechanism. This knowledge-guided design improves reservoir-type discrimination, especially for classes with similar log responses in complex oil–water intervals.

(4) This study demonstrates that knowledge-enhanced attention networks enable not only high-precision permeability prediction but also reliable sweet spot identification in tight sandstone reservoirs. The proposed permeability-driven, cascaded framework establishes a genuinely integrated workflow for reservoir parameter estimation and fluid identification, providing a feasible approach for fine-scale reservoir evaluation and offering potential for extension to three-dimensional modeling and cross-block applications. Overall, the proposed workflow has clear industrial potential for rapid reservoir screening and well-placement optimization in tight sandstone development, particularly when core data are limited.

Author contributions

**[Xinyu Li]: ** Conceptualization, Methodology, Writing—original draft, Writing—review and editing; **[Yuming Liu]: ** Writing -review & editing; **[BingBing Zhang]: ** Formal analysis; **[JingJing Luo]: ** Investigation; **[Hengzhi Liu]: ** data curation; **[Qi Chen]: ** Validation, Visualization. All authors have read and agreed to the published version of the manuscript.

Funding

This research was funded by National Natural Science Foundation of China (No. 42172154; No. 42472205).

Data availability

For confidentiality reasons, some of the data in the article cannot be publicly displayed. If you have data-related questions, you are welcome to contact me by email.

Declarations

Competing interests

The authors declare no competing interests.