A method, medium, device and product for predicting oil saturation
By employing two-stage training and multi-task learning of the TabPFN model, combined with physics-driven synthetic data generation and uncertainty calibration, the problems of data scarcity and overfitting in oil saturation prediction by machine learning models are solved, achieving efficient and reliable prediction results and supporting engineering decision-making.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA UNIV OF GEOSCIENCES (WUHAN)
- Filing Date
- 2026-01-15
- Publication Date
- 2026-06-02
Smart Images
Figure CN122133430A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of oil saturation prediction technology, and in particular to a method, medium, equipment and product for predicting oil saturation. Background Technology
[0002] Traditional methods for predicting oil saturation, such as the Archie equation, rely on empirical parameters like the cementation index and saturation index, which often suffer from insufficient accuracy in argillaceous sandstone or fractured reservoirs due to nonlinear relationships. The application of machine learning in petroleum engineering began in the late 1990s, with researchers exploring artificial intelligence to address the limitations of traditional reservoir characterization methods. With increased computing power and the accumulation of big data, machine learning has evolved from early applications of artificial neural networks (ANNs) (such as the research by Al-Marhoun and Osman in 2002) to deep learning models. After 2010, with the popularization of the concept of smart oilfields, machine learning has been widely used to integrate wellhead, seismic, and core data to predict oil saturation, supporting reserve estimation, waterdrive efficiency calculation, and capillary pressure derivation. This background reflects the shift in the petroleum industry from a physical model-driven to a data-driven paradigm, especially in complex geological environments where machine learning provides more flexible solutions. According to the literature review, early studies such as Anifowose (2012) emphasized the role of machine learning in reservoir characterization, while in recent years, hybrid models such as Sun et al. (2019) have further incorporated physical laws to improve the robustness of predictions.
[0003] Machine learning methods primarily include supervised learning, semi-supervised learning, and deep learning frameworks for predicting oil or water saturation from wellhead data (oil saturation is typically calculated by subtracting water saturation from 1). Common input features include depth (DEPT), neutron porosity (HTNP), true resistivity (RT), gamma rays (CGR / SGR), sonic travel time (DTCO), and volumetric density (RHOZ). For example, Support Vector Machines (SVM) handle nonlinear data with radial basis function kernels. Random Forests (RF) and Least Squares Boosting (LSBoost) serve as ensemble methods to reduce overfitting risk, while Bayesian models provide uncertainty quantification. Deep learning surrogate models, such as those using recurrent neural networks (RNNs) and convolutional neural networks (CNNs), predict oil, water, and gas saturation distributions for water injection processes in monolayer heterogeneous reservoirs. Hybrid methods, such as Physical Information Neural Networks (PINNs) embedding physical laws (e.g., Arps decay models), perform well in spatiotemporal prediction. Furthermore, the combination of surrogate models and numerical simulations allows for rapid assessment of saturation changes, supporting carbon capture and enhanced recovery (EOR) applications.
[0004] Despite progress, major issues still constrain further application. First, data scarcity and quality are common challenges: wellhead data often contains noise, missing values, and outliers; anomaly removal assuming a normal distribution may lose marginal information, leading to model bias. Second, overfitting is risky, especially on small datasets, requiring early stopping, regularization, and cross-validation for mitigation, but models like RNNs are susceptible to gradient vanishing. Third, the black-box nature of the model results in a lack of interpretability, making it difficult for users to understand the predictive basis; while tools like SHAP or LIME can assist, they are not widely integrated. Fourth, generalization ability is limited: models trained on specific oilfields are difficult to directly apply to carbonate or fracture systems, requiring retraining or transfer learning. Fifth, computational resource requirements are high; deep model training is time-consuming, necessitating lightweight architectures for real-time applications. Furthermore, uncertainty propagation is not adequately addressed, and resolution of dynamic reservoir behavior (such as thin-layer effects) is insufficient. Sixth, the model lacks physical priors and uncertainty metrics, making it difficult to support engineering decisions. Summary of the Invention
[0005] The purpose of this invention is to address the problem that existing models lack physical priors and are therefore unable to support engineering decisions. A method for predicting oil saturation is proposed, comprising the following steps: S1. Collect the original logging curves, the discrete auxiliary information corresponding to the logging curves, and the core images, and obtain the oil saturation measured by the laboratory core. Perform data preprocessing to construct the feature X and the oil saturation label vector y. S2. Simulate the relationship between oil saturation and logging response under different formation types, fracture development degrees and oil-bearing conditions, generate logging responses and add noise to obtain noisy synthetic data, including synthetic feature X' and synthetic oil saturation label vector y'. S3. Construct an oil saturation prediction model based on the TabPFN framework, pre-train the model using noisy synthetic data, and fine-tune the pre-trained model using the feature matrix X and label vector y. S4. When the uncertainty of the predicted value of the fine-tuned model exceeds the set threshold, the oil saturation calculated by the Archie physical model is used as the safe output; otherwise, the oil saturation is predicted using the fine-tuned model.
[0006] Furthermore, the logging curves collected in S1 include: natural gamma ray (GR), spontaneous potential (SP), borehole diameter (CAL), density (DEN), sonic transit time (AC), formation true resistivity (RT), and formation resistivity (RXO) of the flushed zone. Discrete auxiliary information includes: well number, block, wellbore structure, and logging tool model.
[0007] Furthermore, data preprocessing includes: Depth interpolation is used to achieve uniform depth alignment for all logging curves; local z-score outlier detection is performed on each curve, and outliers are replaced with the median of the curve value; a missing mask is generated for missing parts of the curve; the logging curve data is obtained after the above processing. ; By performing sliding window statistics on the curve at multiple scales, multiple statistical features are extracted, resulting in a multi-scale statistical feature set. ; Perform wavelet or short-time Fourier transform on the signal of the curve to obtain its frequency domain characteristics. ; Multidimensional vectors are extracted from core images using a convolutional model. ; Obtain the embedding vector e of the discrete auxiliary information; Constructing the eigenvector: d represents the logging depth.
[0008] Furthermore, the relationship between the true resistivity RT of the formation and the oil saturation is as follows:
[0009]
[0010]
[0011] in, This represents the true resistivity of the formation, and 'a' represents the lithology coefficient. Indicates the resistivity of formation water. The value represents porosity, m represents the cementation index, and n represents the saturation index. Indicates water saturation. Indicates oil saturation; The logging response with added noise is:
[0012] in, This represents the true resistivity of the formation after noise has been added. This is tool noise.
[0013] Furthermore, in S4, the threshold is set as follows:
[0014] in, This indicates the set oil saturation threshold. This represents the oil saturation calculated using the Archie physical model.
[0015] Furthermore, during pre-training and fine-tuning, the model employs a multi-task learning framework, with oil saturation prediction as the primary task and auxiliary tasks of porosity and / or permeability prediction simultaneously. The training loss function is:
[0016]
[0017] Where L is the training loss function, , and These are the weighting coefficients for predicting oil saturation, porosity, and permeability, respectively. This represents the loss function for the auxiliary task. This represents the weight coefficient of the regularization term. This represents the regularization loss function used to constrain the size of the model parameters;
[0018] in, This represents the variance of the training sample x. Let x represent the mean of the training samples, and y represent the label.
[0019] The present invention also proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described oil saturation prediction method.
[0020] The present invention also proposes an electronic device, including a processor and a memory, wherein the processor and the memory are interconnected, the memory is used to store a computer program, the computer program including computer-readable instructions, and the processor is configured to invoke the computer-readable instructions to execute the above-described oil saturation prediction method.
[0021] The present invention also proposes a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the above-described oil saturation prediction method.
[0022] The beneficial effects of the technical solution provided by this invention are: (1) This invention obtains multimodal logging curve data by depth alignment, outlier replacement, construction of missing mask, calculation of window statistical features and construction of curve frequency domain features, and extracts feature vectors of core images and embedding vectors of scattered data. All data constitute the input feature vector, which solves the problems of asynchronous logging data, high noise, severe missing data and weak multi-scale structure.
[0023] (2) This invention simulates the relationship between oil saturation and logging response through physical experience, and constructs a physical-driven synthetic sample generator to solve the problems of scarcity of real label samples, high sampling cost and weak cross-well generalization.
[0024] (3) This invention uses the TabPFN model for two-stage training: First, prior pre-training is performed on synthetic samples to enable the model to learn the basic mapping relationship between logging curves and saturation, and to obtain a general prior distribution structure across wells; then, the model is loaded into real data for fine-tuning with a small learning rate, and the deviation of the physical model is further corrected through real labels to make it more adaptable to the actual logging response of the target oil and gas field. This solves the problem of severe overfitting of traditional machine learning methods in small sample and cross-well environments, and gives the model the advantages of strong prior, fast fine-tuning, and good generalization, making it easy to support engineering decisions. Attached Figure Description
[0025] Figure 1 This is a flowchart of the oil saturation prediction method according to an embodiment of the present invention; Figure 2 This is a graph showing the fitting effect of the prediction results of this invention; Figure 3 This is a diagram showing the predicted oil saturation effect of well HD1-13H according to the present invention. Figure 4 This is a diagram showing the predicted oil saturation effect of well HD1-27H according to the present invention. Figure 5 This is a block diagram of an electronic device according to an exemplary embodiment of the present invention. Detailed Implementation
[0026] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described below with reference to the accompanying drawings.
[0027] The flowchart of the oil saturation prediction method of this invention is as follows: Figure 1 Specifically, it includes the following steps: S1. Collect the original logging curves, the discrete auxiliary information corresponding to the logging curves, and the core images. Obtain the oil saturation measured by the laboratory core and perform data preprocessing to construct the feature X and the oil saturation label vector y.
[0028] The original logging curves include: natural gamma ray (GR), spontaneous potential (SP), wellbore diameter (CAL), density (DEN), sonic transit time (AC), formation true resistivity (RT), and formation resistivity (RXO) of the flushed zone. Discrete auxiliary information corresponding to the logging curves is also collected, such as: well number, block, wellbore structure, and logging tool model. Core photographs / thin section images are also included.
[0029] The above data is preprocessed, including: (1) Achieve uniform depth alignment of all logging curves through depth interpolation; perform local z-score outlier detection on each curve and replace outliers with the median of the curve values; generate a binary missing mask for the missing part of the curve for each depth point; obtain logging curve data after the above processing. .
[0030] (2) Perform sliding window statistics on the curve at each depth point at multiple scales, extract multiple statistical features, and obtain a multi-scale statistical feature set. .
[0031] (3) Perform wavelet or short-time Fourier transform on the signal of the curve at each depth point to obtain the frequency domain features. .
[0032] (4) Use a pre-trained convolutional model to extract 128 or 256-dimensional vectors from the core images. .
[0033] (5) Obtain the embedding vector e of discrete data (well number, block, well structure, and logging tool model); Constructing the eigenvector: d represents the logging depth.
[0034] Label (If there is no label, it is recorded as an unlabeled sample).
[0035] S2. Simulate the relationship between oil saturation and logging response under different formation types, fracture development levels, and oil-bearing conditions. Generate logging responses and add noise to obtain noisy synthetic data, including synthetic features X' and synthetic oil saturation label vector y'.
[0036] To address the issues of scarce real-label samples, high sampling costs, and weak cross-well generalization, this invention constructs a physical-driven synthetic sample generator to produce a large number of "conditional" training samples consistent with geological patterns.
[0037] First, the distribution of basic physical properties such as porosity, specific surface area, and fracture grade are defined for different lithological types. Then, the permeability is calculated based on the sampled porosity and specific surface area using the Kozeny–Carman equation, and water saturation and oil saturation are generated according to the water-bearing conditions.
[0038] These physical property values are input into the Archie resistivity model, and controllable noise and offset are added according to the logging tool response to simulate the physical logging response of multiple logging curves such as RT, DEN, AC, and GR.
[0039] The relationship between formation true resistivity RT and oil saturation in well logging response is as follows:
[0040]
[0041]
[0042] in, This represents the true resistivity of the formation, and 'a' represents the lithology coefficient. Indicates the resistivity of formation water. The value represents porosity, m represents the cementation index, and n represents the saturation index. Indicates water saturation. This indicates the degree of oil saturation.
[0043] The logging response with added noise is:
[0044] in, This represents the true resistivity of the formation after noise has been added. This is tool noise.
[0045] To simulate the missing patterns of different logging tools in real wells, the system further randomly drops tracks within local depth ranges according to probability, thereby generating a missing mask structure that is completely consistent with the real logging data, making the synthetic data closer to the real data in terms of statistical distribution structure. The synthetic data generated in this step is consistent with the real data formed in step S1 in terms of feature structure.
[0046] S3. Construct an oil saturation prediction model based on the TabPFN framework, pre-train the model using noisy synthetic data, and fine-tune the pre-trained model using the feature matrix X and label vector y.
[0047] After obtaining unified feature vectors from both real and synthetic data, this invention uses the TabPFN model for two-stage training: First, prior pre-training is performed on a large number of synthetic samples, enabling the model to learn the basic mapping relationship between logging curves and saturation on a large scale of "physically consistent" samples that do not require manual measurement, thus obtaining a universal prior distribution structure across wells; then, the model is loaded into real data for fine-tuning with a small learning rate, and the deviations of the physical model are further corrected through real labels, making it more adaptable to the actual logging response of the target oil and gas field. During the fine-tuning stage, the TabPFN model can quickly adjust its attention mechanism according to the data distribution structure of real wells, ensuring that the model maintains stable performance across different wells (cross-well migration).
[0048] During the two-stage training, a multi-task learning framework was adopted, with oil saturation prediction as the primary task and porosity and / or permeability prediction as secondary tasks. The training loss function was:
[0049]
[0050] Where L is the training loss function, , and These are the weighting coefficients for predicting oil saturation, porosity, and permeability, respectively. This represents the loss function for the auxiliary task. This represents the weight coefficient of the regularization term. This represents the regularization loss function used to constrain the size of the model parameters.
[0051]
[0052] in, This represents the variance of the training sample x. Let x represent the mean of the training samples, and y represent the label.
[0053] S4. When the uncertainty of the predicted value of the fine-tuned model exceeds the set threshold, the oil saturation calculated by the Archie physical model is used as the safe output; otherwise, the oil saturation is predicted using the fine-tuned model.
[0054] To ensure the reliability of oil saturation predictions in practical engineering scenarios, this invention constructs a complete uncertainty measurement, calibration, and backoff system. During inference, the model outputs the predicted mean and variance, while simultaneously estimating input anomalies through multi-model ensemble and feature space distance. When the model's output variance is too large or the input deviates from the training distribution, the system automatically triggers a confidence region determination and performs variance calibration: a variance scaling factor is learned in the validation set to achieve target coverage (e.g., 95%) in the prediction interval. If the calibrated uncertainty still exceeds a threshold, the system automatically initiates a backoff strategy, using the baseline oil saturation calculated by the Archie physical model as a safety output to avoid engineering misjudgments due to insufficient model confidence.
[0055] The threshold set in this invention is:
[0056] in, This indicates the set oil saturation threshold. This represents the oil saturation calculated using the Archie physical model.
[0057] To verify the effectiveness of the method of the present invention, several evaluation indicators were constructed, including MAE (mean absolute error), RMSE (root mean square error), and R² (coefficient of determination). Figure 2 This is the fitting effect of the prediction results of this invention. Figure 2In this study, the coefficient of determination for the prediction results obtained by the method of this invention is 0.994, and the root mean square error is 1.5. From... Figure 2 As can be seen, after adding pre-training, the fitting points of this invention converge significantly, getting closer to the ideal 1:1 straight line, and the model error is significantly reduced.
[0058] Figure 3 and Figure 4 These are the oil saturation predictions for wells HD1-13H and HD1-27H, as predicted by this invention. It can be seen that the prediction curves output by the method of this invention are smooth and have a reasonable trend. When encountering abnormal depth segments, the prediction interval automatically widens, demonstrating the ability of the uncertainty mechanism to mitigate engineering risks.
[0059] In one exemplary embodiment, a computer-readable storage medium is included, which stores a computer program that, when executed by a processor, implements the oil saturation prediction method described above.
[0060] Please see Figure 5 In one exemplary embodiment, the device further includes an electronic device including at least one processor, at least one memory, and at least one communication bus.
[0061] The memory stores a computer program, which includes computer-readable instructions. The processor calls the computer-readable instructions stored in the memory through the communication bus to execute the oil saturation prediction method described above.
[0062] In one exemplary embodiment, a computer program product is proposed, including a computer program / instructions that, when executed by a processor, implement the steps of the oil saturation prediction method described above.
[0063] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A method for predicting oil saturation, characterized in that, Includes the following steps: S1. Collect the original logging curves, the discrete auxiliary information corresponding to the logging curves, and the core images, and obtain the oil saturation measured by the laboratory core. Perform data preprocessing to construct the feature X and the oil saturation label vector y. S2. Based on physical empirical formulas, simulate the relationship between oil saturation and logging response under different formation types, fracture development degrees and oil-bearing conditions, generate logging responses and add noise to obtain noisy synthetic data, including synthetic feature X' and synthetic oil saturation label vector y'. S3. Construct an oil saturation prediction model based on the TabPFN framework, pre-train the model using noisy synthetic data, and fine-tune the pre-trained model using the feature matrix X and label vector y. S4. When the uncertainty of the predicted value of the fine-tuned model exceeds the set threshold, the oil saturation calculated by the Archie physical model is used as the safe output; otherwise, the oil saturation is predicted using the fine-tuned model.
2. The method for predicting oil saturation according to claim 1, characterized in that, The logging curves collected in S1 include: natural gamma ray (GR), spontaneous potential (SP), borehole diameter (CAL), density (DEN), sonic transit time (AC), formation true resistivity (RT), and formation resistivity (RXO) of the flushed zone. Discrete auxiliary information includes: well number, block, wellbore structure, and logging tool model.
3. The method for predicting oil saturation according to claim 1, characterized in that, Data preprocessing includes: Depth interpolation is used to achieve uniform depth alignment for all logging curves; local z-score outlier detection is performed on each curve, and outliers are replaced with the median of the curve value; a missing mask is generated for missing parts of the curve; the logging curve data is obtained after the above processing. ; By performing sliding window statistics on the curve at multiple scales, multiple statistical features are extracted, resulting in a multi-scale statistical feature set. ; Perform wavelet or short-time Fourier transform on the signal of the curve to obtain its frequency domain characteristics. ; Multidimensional vectors are extracted from core images using a convolutional model. ; Obtain the embedding vector e of the discrete auxiliary information; Constructing the eigenvector: d represents the logging depth.
4. The method for predicting oil saturation according to claim 1, characterized in that, The relationship between formation true resistivity RT and oil saturation is as follows: in, This represents the true resistivity of the formation, and 'a' represents the lithology coefficient. Indicates the resistivity of formation water. The value represents porosity, m represents the cementation index, and n represents the saturation index. Indicates water saturation. Indicates oil saturation; The logging response with added noise is: in, This represents the true resistivity of the formation after noise has been added. This is tool noise.
5. The method for predicting oil saturation according to claim 1, characterized in that, In S4, the threshold is set as follows: in, This indicates the set oil saturation threshold. This represents the oil saturation calculated using the Archie physical model.
6. The method for predicting oil saturation according to claim 1, characterized in that, During pre-training and fine-tuning, the model employs a multi-task learning framework, with oil saturation prediction as the primary task and auxiliary tasks of porosity and / or permeability prediction as the secondary task. The training loss function is: Where L is the training loss function, , and These are the weighting coefficients for predicting oil saturation, porosity, and permeability, respectively. This represents the loss function for the auxiliary task. This represents the weight coefficient of the regularization term. This represents the regularization loss function used to constrain the size of the model parameters; in, This represents the variance of the training sample x. Let x represent the mean of the training samples, and y represent the label.
7. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, it implements the method as described in any one of claims 1-6.
8. An electronic device, characterized in that, The device includes a processor and a memory, the processor being interconnected with the memory, wherein the memory is used to store a computer program, the computer program including computer-readable instructions, and the processor is configured to invoke the computer-readable instructions to perform the method as described in any one of claims 1-6.
9. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1-6.