4.1. Data Description and Processing
A total of 12 core wells, designated from A to L, were selected for training in this study. The details of the dataset are presented in
Table 1. Wells A through K were utilized as the training and validation sets, while well L was employed as a test set to assess the model’s generalization ability. Firstly, the sedimentary facies at the corresponding depths were manually identified for 12 core wells by observing the core photos as well as checking the textual descriptions as shown in
Table 2, where the sedimentary microfacies codes are shown in
Table 3, and the manually identified microfacies were used as the labeling data for the training model.
(1) A total of 16 curves, including SP (spontaneous potential), AC (acoustic), RT (true formation resistivity), and RI (resistivity index) were selected for the logging curve dataset. The GR curve (among logging curves) is typically employed to reflect the lithological change characteristics. However, the region is subject to volcanic activity (
Figure 7 illustrates the tuff of well L), and the GR curve exhibits outliers, rendering it unsuitable for analysis [
55,
56].
(2) The physical property curves include ART, ARI, AAC, POR, PERM, SO, and others. Among these, the POR, PERM, and SO curves provide a more intuitive reflection of the porosity, permeability, and fluid content of the rocks. These curves are crucial for highlighting the characteristics of sedimentary facies and aiding in the identification of stratigraphic properties. For example, higher porosity and permeability are commonly found in sandy sedimentary facies with high storage capacity, such as distributary channels and debouch bars.
(3) The format of the data pertaining to both the logging curve and the physical curve is standardized, obviating the necessity for extensive processing, with the exception of the removal of outliers. However, the text description data and core image data necessitate a more intricate data preprocessing procedure. The text description data must first be extracted in order to obtain key features such as top depth, bottom depth, color, grain size, composition, lithology, and so forth. This is necessary to ensure that the extracted information is accurate and structured. Subsequently, the extracted data must be cleaned to remove invalid or anomalous values, address missing data, and standardize the data format, unify the description of color, grain size, composition, and ensure the consistency of description between different records. Subsequently, the category-based features are converted to numerical features via label encoding (LEC), thereby enabling the model to process the features.
(4) The core images require a lot of manual labeling, and all the core images need to be labeled with depth and sedimentary facies information, and the labeling rule is “well name_cylinder number-section number_core top depth_core bottom depth_sedimentary microfacies”. After labeling, all the images were enhanced and preprocessed: firstly, the image size was adjusted to 224 × 224 pixels. Then random horizontal flipping and rotation were applied to enrich the image data. After that, the image data were converted into a tensor and normalized by using ImageNet’s standard mean and standard deviation.
The multimodal data used in this study, including logging curves, physical properties, and core images, exhibited significant heterogeneity due to differences in their measurement units and data acquisition methods. To address the challenges posed by the heterogeneity of the data, we employed several preprocessing techniques to minimize potential systematic errors. These techniques included data normalization to standardize measurement scales across different data types and noise reduction methods to eliminate spurious data points. Furthermore, we applied cross-validation to assess the model’s performance across multiple folds, ensuring that the model remains robust even when trained on varied and heterogeneous datasets. These measures were crucial in minimizing the risk of systematic errors arising from data heterogeneity, thereby improving the model’s reliability and generalizability.
Once the data were subjected to the requisite preprocessing, the sedimentary facies were correlated with the logging data, the physical property data, and the text descriptions, according to depth. The core image data were manually labeled and, thus, did not require further processing. Ultimately, the data from well L were stored as part of a distinct test set, while the remaining data were employed in the training set and validation set for model training.
4.2. Unimodal Experiments and Results
4.2.1. Well Logs Prediction Experiment
The computations for this experiment and the following experiments were conducted on a system equipped with an Intel Core i5-12400F CPU at 2.50 GHz, supported by 32 GB RAM. The device operated on a Win10 64-bit OS and featured an NVIDIA GeForce RTX 4060ti graphics card with 16 GB memory.
In order to predict sedimentary facies, the random forest algorithm was selected as the most appropriate for the logging data. Once the preprocessing of the logging data was completed, the optimal hyperparameters for the random forest were selected using the method of cross-validation.
In order to find the optimal hyperparameter combinations, the range of each hyperparameter was first defined: number of trees: 100, 200, 300, 400, 500; maximum depth: 10, 20, 30, 40, 50; the maximum number of features: ‘auto’, ‘sqrt’, ‘log2’; minimum number of sample splits: 2, 5, 10; and minimum number of sample leaves: 1, 2, 4. Afterward, all of the possible hyperparameter combinations were used to evaluate the model performance using a 5-fold cross-validation approach. In each fold cross-validation, the dataset was randomly partitioned into five subsets, where four subsets were used for training and one subset was used for validation. This process was repeated five times. Each time, a different subset was selected as the validation set, and the other four subsets were used as the training set. Based on the results of cross-validation, the combination of hyperparameters with the highest average accuracy was selected and the final random forest model was trained using the best hyperparameters. After the model was trained, the performance of the model was evaluated using metrics such as accuracy, precision, recall, and F1 score, as shown in
Table 4. The model is presented with a confusion matrix as shown in
Figure 8a.
It is obvious from the confusion matrix that the overall prediction accuracy of the model is as high as 94%. The best prediction is in the main channel (Facies 1) and interchannel (Facies 4), with a recall as high as 97%, while the prediction is poorer in the channel edge (Facies 3) and debouch bar (Facies 5), with a recall lower than 85%, which is partly due to the difference in the number of datasets, with the number of samples in the channel edge and debouch bar significantly less than the other three microfacies, which leads to their lower recall along with higher precisions.
On the other hand, the main channel is usually characterized by high-energy sedimentary features, such as coarser grains and the result of higher porosity, while the interchannel is characterized by finer grains, lower porosity, and more muddy constituents; these significant differences in the features allow the model to more accurately differentiate between these two types of microfacies.
It can also be seen from the confusion matrix that channel edges are misclassified as the main channel and interchannel, and debouch bars are often misclassified as the distributary channel and interchannel. In addition to the relatively small amount of training data for the channel edge and debouch bar, this misclassification is likely due to the presence of similar lithologies across different sedimentary microfacies. Channel edges feature coarser sedimentary grains and greater thicknesses near the channel, similar to the main channel, while areas farther from the channel are dominated by fine-grained sediments, akin to the interchannel. Debouch bars, which develop between channels, consist mainly of well-sorted fine sand and silt, which are characteristics also common in distributary channels and interchannels.
4.2.2. Physical Properties Prediction Experiment
The physical curves exhibit a data structure analogous to that of the logging curves; however, they are distinguished by a markedly lower degree of complexity and unstructured characteristics, so the relatively simple KNN model is chosen for prediction. The same cross-validation method was selected to adjust the hyperparameters of the KNN model in order to obtain the optimal hyperparameter combination. Ultimately, Manhattan was selected as the metric, 9 for K, and distance for weighted voting weights. The final model performance is presented in
Table 5, and the confusion matrix is shown in
Figure 8b.
From the model output results, the accuracy of predicting sedimentary microfacies from physical property data is significantly lower than that of logging curves, which may be attributed to the small number of physical property curves and the limitation of the physical properties they reflect, which cannot comprehensively capture the complex characteristics of sedimentary microfacies. In addition, the characteristic differences between different sedimentary microfacies in the physical property curves may not be significant enough to cause the model to perform poorly in distinguishing between different sedimentary microfacies. From the confusion matrix, the recalls of the channel edge and debouch bar are lower than 0.5, in which the probability of the channel edge being misidentified as the main channel, distributary channel, and interchannel is not much different, which may be due to the small number of samples on the one hand. On the other hand, it may be due to the complexity of the lithology of the channel edge, the lithology is similar to the three microfacies, such as the main channel, resulting in physical characteristics that cannot be well distinguished from the three microfacies. The training results for the debouch bar may be suboptimal due to the limited amount of data, which makes it challenging to accurately distinguish these sedimentary microfacies from others.
4.2.3. Textual Descriptions Prediction Experiment
The textual description dataset has a total of 474 data points following data preprocessing, which is a considerably smaller number than that of the other modal dataset. In the event of insufficient data, a single classifier may result in overfitting of the training data, whereas the bagging classifier can enhance the model’s generalizability by integrating the prediction outcomes of multiple classifiers. The integrated learning feature of the bagged classifier can compensate for the shortcomings brought about by the small data size to some extent, thus improving the overall performance and stability of the model. Again, after hyperparameter selection, the model is trained and the performance of the model is evaluated on the validation set as shown in
Table 6, and the confusion matrix is shown in
Figure 8c.
From the model outputs, the overall prediction of the textual description data is biased due to the small sample data size, especially at the river edges and debouch bar, where the prediction accuracies and recalls are low due to the significantly too small data size. Furthermore, the confusion matrix indicates that the distributary channel, channel edge, and interchannel were erroneously identified as the main channel. This likely resulted from the higher number of main channel samples compared to the other microfacies samples, which biased the model toward identifying the main channel during training. The limited number of samples presents a challenge for the model in accurately discerning the distinctive characteristics of the other microfacies. However, the textual description data themselves are more sparse and have an accuracy of more than 70%, which is not as precise as the models trained on other data but can contribute to the overall prediction of sedimentary microfacies.
4.2.4. Core Image Prediction Experiments
The classic convolutional neural network model was selected for the core image prediction experiments for training. The total number of pre-processed core photos was 530, which may be considered a relatively modest amount of data for a complex classification task such as predicting sedimentary microfacies. Accordingly, the ResNet-50 pre-trained model was selected as the base model; ResNet-50 was pre-trained and fine-tuned, with a fully connected layer added on top to output sedimentary facies classification results. The results of the model training are presented in
Table 7, and the confusion matrix is shown in
Figure 8d.
From the table, it can be seen that the prediction results are poor due to the small number of core image samples in the channel edge and debouch bar, whereas the prediction results for the main channel, distributary channel, and interchannel are better, with prediction accuracies of about 90% and above. Since the same type of lithology may correspond to more than one sedimentary microfacies, the use of images to predict sedimentary microfacies is subject to a certain degree of misclassification, e.g., fine sandstones may be present in both the interchannel and debouch bar. This is another indication that the accuracy and generalizability of models using only unimodal data for sedimentary facies prediction are prone to bottlenecks.
4.3. Multimodal Data Fusion Model
After training the four models, realizing decision-level multimodal fusion requires further processing of the data. First, the logging data and physical data are organized in a one-to-one correspondence according to depth. That is to say, the logging data and physical data of the same depth are placed in the same row. As the depth of each textual description sample and core image sample is not a single depth point, but rather a section of depth, each row represents a depth interval. In order to align with the logging and physical data, it is necessary to correspond each depth point to the depth interval in which it is located.
That is, if a depth point from a specific set of logging and physical properties data falls within the depth interval of a particular piece of textual description data, this textual description data is assigned to that depth point. Similarly, if a depth point is situated within the depth interval of a core image, that core image is assigned to that depth point. However, there may be cases where a depth point has only logging and physical data, with no corresponding textual description or core image; these data points will be removed.
Following the aforementioned processing, data with disparate depth resolutions can be aligned and integrated into a comprehensive dataset. This alignment and integration guarantees that each depth point contains multimodal data from which features can be extracted for multimodal data fusion prediction. Although this approach introduces additional complexity, it can leverage the distinctive attributes of each data type, thereby enhancing the model’s predictive accuracy.
As mentioned in the previous section, commonly used methods in multimodal data fusion research include feature-level fusion and decision-level fusion. Feature-level fusion can be defined as the integration of features derived from individual modal data, which have been extracted and subsequently fed into a unified model for classification or regression purposes. The benefit of this methodology is that it allows for the comprehensive exploitation of the granular data pertaining to all modal variables, thereby enhancing the overall performance of the model. However, a significant drawback of feature-level fusion is that when the data dimensions and features of each modality differ significantly, the fused feature space may become too complex, thereby rendering the model challenging to train. It is possible that if only a limited number of data types are subjected to training, satisfactory outcomes may be attained. However, there are a plethora of information modalities in the oilfield. While this paper mentions only four modalities, for a study area containing a greater number of modal data, employing feature-level fusion is not conducive to incorporating additional modalities into the original dataset. Moreover, feature-level fusion requires extensive data preprocessing, which involves standardizing and aligning data from different modalities. This introduces additional complexity when dealing with a large amount of heterogeneous data, such as oilfield data, and significantly increases the difficulty of data preprocessing.
In contrast, decision-level fusion integrates the prediction outcomes of each model after the processing and prediction of the respective modal data by discrete models. This is typically achieved through weighted voting or alternative aggregation techniques, thereby yielding the ultimate decision. The advantage of decision-level fusion is that it preserves the independence of each modality and is more flexible in dealing with heterogeneous data from multiple sources. In comparison to feature-level fusion, decision-level fusion methods are relatively straightforward, do not require complex preprocessing and standardization of data from different modalities, and are well-suited to the processing of large-scale and complex datasets.
It can be concluded that the decision-level fusion method is a more suitable approach for analyzing the data in the oilfield context. Firstly, suitable models are selected for training and feature extraction for each of the four modalities. Subsequently, the prediction results of each model are weighted according to the assigned weights, which are based on the performance of each model on the validation set. The weights are 0.94, 0.76, 0.72, and 0.90, respectively. Finally, the prediction results of each model are weighted and summed up, and the category with the highest score is selected as the final prediction result; the results are shown in
Table 8 and
Figure 9a.
The final prediction results indicate a significant enhancement in the efficacy of multimodal data fusion, accompanied by a notable improvement in the recognition rates of the remaining four sedimentary microphases, with the exception of the debouch bar, which exhibited relatively lower accuracy. The performance and effect of the multimodal data fusion model are observed to be markedly superior to those of the unimodal data prediction model. For the identification of the sedimentary phase studied in this paper, the formation mechanism is complex, often requiring the consideration of multiple factors. It is clear that relying solely on unimodal prediction, or considering only one type of factor, yields insufficient results; the accuracy of predictions using the physical property curve and text description does not exceed 80%. Even if the prediction results are good, there may be some randomness. However, multimodal prediction can synthesize various elements. When modal data are missing or the prediction effect is poor, data from other modes can compensate for the shortcomings of the entire network, significantly enhancing the credibility of the prediction results for complex data compared to unimodal prediction.
The data from the test set of well L are predicted by the model after decision-level fusion; the results are shown in
Table 9 and
Figure 9b. The overall accuracy reaches 93%, which is very close to that of the validation set, further confirming the practical application of the fusion model.
To further verify the feasibility of the method in sedimentary facies identification application, we selected four wells in a paragenetic profile for sedimentary facies prediction. Among these four wells, three wells on the left were non-core wells containing only the logging curve and physical property curve data, and the other well was a core well containing all data. The effect of the data presentation after prediction is shown in
Figure 10, which is basically consistent with the recognition of the study area [
53]. It can be seen that the method can be applied to the whole study area to predict and recognize the sedimentary phase for all wells and construct the sedimentary grid of the study area.