Before constructing the massive-scale benchmark dataset of seismic facies, it is necessary to develop a knowledge graph of seismic facies based on the geological concepts and seismic reflection configurations, which can provide guidelines for preparing representative dataset samples and assessing facies classification results. Based on specific seismic reflection configurations, seismic facies can be roughly divided into parallel and subparallel, prograding clinoforms, fill, hummocky, chaotic, divergent, wave, and reflection-free (Fig. ). These seismic facies can be subdivided further based on several independent parameters, such as reflection configurations, continuity, amplitude, and frequency. For example, parallel and subparallel reflections can be subdivided into 27 different types based on frequency (high, middle, and low), amplitude (strong, moderate, and week) and continuity (excellent, medium, and poor). Based on different reflection patterns, prograding clinoforms, fill, and hummocky can be subdivided further into five (sigmoid, oblique, shingled, parallel, and complex), six (onlap, prograded, mounded onlap, divergent, chaotic, and complex), and four (fan complex, mound, blanking, and chaotic) types, respectively.

As shown in Fig. , we develop a knowledge graph of seismic facies and illustrate the typical seismic reflection configurations for eight types of seismic facies. However, considering the requirements for data amount and diversity in this work, we take the five most common seismic facies (parallel and subparallel, prograding clinoforms, fill, hummocky, and chaotic) as examples to explain how to construct a massive-scale, feature-rich, and high-realism benchmark dataset of seismic facies from field data curation, knowledge-guided synthesization, and GAN-based generation.

We start building our benchmark dataset by employing the field seismic data curation strategy with a series of steps, i.e., raw data collection, manual interpretation and classification, bandpass filtering, resampling, amplitude equalization, and skeletonization. We first collect almost 4000 global publicly available 2-D seismic profiles and 10 3-D seismic data from the United States Geological Survey (USGS), New Zealand Petroleum & Minerals (NZPAM), the South Australian Resources Information Gateway (SARIG), and the Society of Exploration Geophysicists (SEG). These 2-D and 3-D seismic data amount to around 130 G and are primarily located in the Gulf of Mexico, on the eastern and western coasts of the United States, in Alaska, in the Bering Sea, in the Beaufort Sea, in New Zealand, in South Australia, and in the Sichuan Basin (see the data distribution map in Fig. ).

We then manually select, crop, and classify these field seismic data based on the knowledge graph (Fig. ). As shown in the raw seismic data of Fig. , in total we collect 1000, 700, 500, 500, and 700 2-D raw seismic data for five common seismic facies. However, due to the different data sources, depositional environments, and data processing methods, these raw seismic data have large differences in their sampling rates, amplitudes, and frequency distributions (as shown in Figs.  and a) among the same and different classes of seismic facies. These data variations and uncertainties are not related to the seismic facies. Moreover, they may pose significant obstacles to deep-learning models when learning crucial features such as texture patterns and reflection configurations, which are essential for identifying seismic facies categories. To eliminate such uncertainties in building our standard benchmark dataset, we introduce the data standardization process (Fig. ) for each raw seismic datum, including filtering, resampling, amplitude equalization, and frequency equalization. After applying the data standardization process, the processed seismic data are significantly improved in the consistency of the sampling rates, amplitudes, and frequency distributions (as shown in Figs. b and ). Finally, we retain the main geological structure information of strata by keeping only the waveform peaks as ones and setting zeros elsewhere to obtain the corresponding field skeletonization images shown in Figs. c and .

Compared to the skeletonization images (Fig. d) obtained directly from the raw seismic data, the ones (Fig. c) with data standardization can more clearly reflect the geological structure characteristics and enhance the consistency among the same and different classes of seismic facies. The whole curation strategy, particularly the data standardization processes and skeletonization, eliminates uncertainties inherent in field data from various surveys. This approach only retains the texture patterns associated with seismic facies to produce standardized images for constructing the benchmark dataset. The same processing techniques will also be applied to inference data to ensure that a deep-learning model trained on this dataset achieves consistent predictions.

However, the facies samples from these field seismic datasets only are imbalanced in the categories and lack diversity, and therefore they are not sufficient for building a massive-scale and representative benchmark dataset. For example, parallel and subparallel data are more common than fill or hummocky data in field seismic datasets. Additionally, some specific patterns (e.g., parallel prograding clinoforms, chaotic fill, complex fill, and blanking hummocky) are rare in these publicly available field seismic data.

In order to overcome the sample imbalance and improve the diversity of the dataset, we further develop the second strategy to automatically generate synthetic facies samples based on the knowledge graph of seismic facies and independent seismic reflection configurations. We first define the different geological structural curves by using the following geometric functions: 1z=z0,2z=k0⋅x+z0,3z=k0⋅x2+z0,4z=1k1+k2⋅e-k3⋅x,5z=ek1⋅x-e-k2⋅xek1⋅x+e-k2⋅x, where x and z represent the positions in the crossline and depth directions, respectively. Other parameters (z0, k0, k1, k2, and k3) are used to control the geometry and distribution of the geological structural curves. We then set different combinations of geological structural curves based on different seismic facies categories. Additionally, we randomly set shape parameters for these geological structural curves and combine them at random intervals to enhance their diversity. Furthermore, we first define some key points for some complex geological structures and then generate the corresponding geological structural curves by applying an interpolation process. After generating these different geological structural curves, we add random noise and apply a random local mask to each curve to improve the realism of the synthetic curves. Finally, we set ones on the geological structural curves and zeros elsewhere to generate the corresponding synthetic skeletonization data.

In this way, we randomly generate synthetic facies samples for each type of seismic facies, especially some specific patterns that are rare in field data curation, thus complementing the benchmark dataset of seismic facies. Finally, we automatically generate 500 synthetic facies samples for the five common seismic facies shown in Fig. . Compared to the field facies samples shown in Fig. , the synthetic facies samples generated from knowledge-guided synthesization contain more diverse patterns and reduce sample imbalance. However, these synthetic facies samples may be ideally patterned and lack realism.

As shown in subset-1 and subset-2 in Figs. , , and a, b, the field facies samples have high realism but low diversity, while the synthetic facies samples have high diversity but low realism. In order to construct a comprehensive benchmark dataset of seismic facies, we develop the final strategy of GAN-based generation (Fig. ) to build more facies samples with both high diversity and high realism.

As shown in Fig. c, the architecture of the deep-learning network used in this work is modified from the progressive growing of GANs proposed by . Traditionally, progressive growing of GANs consists of a generator model (G) and a discriminator model (D), where G is used to capture the data distribution and generate fake images to resemble the training dataset (real images) and D is used to assess the probability of images being real or fake. G is composed of a Gen-1 module, five Gen-2 modules, and a Conv_1×1 layer, where the Gen-1 module consists of a 4×4 convolutional layer and a 3×3 convolutional layer and the Gen-2 module consists of an upsampling layer and two 3×3 convolutional layers. D is composed of a Conv_1×1 layer, five Dis-1 modules, and a Dis-2 module, where the Dis-1 module consists of two 3×3 convolutional layers and an average pooling layer and the Dis-2 module consists of a minibatch stddev layer, a 3×3 convolutional layer, a 4×4 convolutional layer, a flattened layer, and a linear layer. Compared to traditional GANs, the progressive growing of GANs does not directly generate high-resolution images but starts by generating simple low-resolution images and then continuously increasing the resolution of the generated images during the network training. This training strategy allows the network to learn the features of the training dataset from coarse to fine scales, resulting in a faster training speed, higher stability, and better-quality images. In addition, we use the Wasserstein GAN with a gradient penalty (WGAN-GP) loss proposed by as the GANs loss function L(G, D) to optimize the network.

We combine subset-1 and subset-2 as training datasets to train the progressive growing of GANs. Initially, we first train a simple network consisting of a Gen-1 module, two Conv_1×1 layers, and a Dis-2 module to generate and access the real and fake facies samples at a 4×4 scale. After stabilizing the training of this simple network, we incorporate a Gen-2 module and a Dis-1 module into it to double the resolutions of G and D. In this way, our network will progressively grow to steadily generate high-resolution (128×128) facies samples. Finally, we use the trained G to automatically generate 500 facies samples for each type of seismic facies shown in subset-3 in Figs. d and . Compared to subset-1 and subset-2, the facies samples constructed by the GAN-based generation have both high diversity and high realism.

After applying the three strategies of field data curation, knowledge-guided synthesization, and GAN-based generation to generate diverse facies samples, we construct a massive-scale, feature-rich, and high-realism benchmark dataset of seismic facies, and we display some facies samples in Figs. –. As shown in Fig. , finally we generate totals of 2000, 1500, 1500, 1500, and 1500 diverse facies samples (128 [inline] × 128 [time]) for five common seismic facies (parallel and subparallel, prograding clinoforms, fill, hummocky, and chaotic). The final benchmark dataset, named cigFacies, has been made publicly available at 10.5281/zenodo.10777460 .

We consider seismic facies classification to be an image classification problem with the goal of classifying 3-D field seismic data according to the corresponding seismic facies (e.g., parallel and subparallel, prograding clinoforms, fill, hummocky, and chaotic). In this study, we use a simple deep-learning network (ResNet-50) proposed by (Fig. a) to implement automatic seismic facies classification. We train and validate our CNN model by using 6400 and 1600 random pairs of facies samples. In addition, in order to improve the diversity of the dataset, we apply random data augmentation strategies (e.g., flip, translation, crop, or resize) for each facies sample before feeding them into the network. We train our network by using the following cross-entropy loss function L: 6L=-∑i=0N-1yilog⁡(xi), where N denotes the number of classes and xi and yi represent the one-hot prediction and label for the ith class. Considering the computation time and memory, we set the batch size to 32 and use the Adam optimizer to optimize the network parameters. In the training process, we start the learning rate at 0.01 and adaptively reduce the learning rate by half when the training metric stagnates within two epochs. As shown in Fig. b, c, both the training loss and validation loss converge to 0.006 and 0.1, while the learning rate decreases from 0.01 to 0.00001 after 200 epochs.

To verify the performance of the trained network, we first apply it to the validation dataset, which is not included in the training dataset. As shown in Fig. d, the predicted results are highly consistent with the labels. In addition, the predicted accuracies for the five common seismic facies in the validation dataset can be up to 97.75%, 99%, 99.67%, 97.33%, and 98.33%, which indicates that the trained network has successfully learned automatic seismic facies classification.

To further verify the performance of the trained model, we develop a prediction workflow shown in Fig. a for automatic seismic facies classification in 3-D field seismic data. We first use an automatic horizon-picking method to extract the top and bottom surfaces (green and red curves in Fig. b) of the target section in the 3-D field seismic data. Then we flatten the field seismic data with the bottom surface to eliminate the influence of the geological structures. In addition, we set a sliding window (blue box in Fig. b) centered on the midpoint of the top and bottom horizons and bounded by these surfaces in order to extract a 2-D raw seismic image pixel by pixel (or trace by trace). The width of the sliding window mainly depends on the size of the classified object within the target section and is typically slightly greater than the widths of the objects. We also apply the standardization and skeletonization processes to the flattened image to make it consistent with the training dataset. Finally, we feed the corresponding skeletonization image into the trained network for automatic seismic facies classification.

In this work, we apply the trained network to two distinct 3-D field seismic data (Longang and Yuanba) with complex geological structures. The Longgang (LG) and Yuanba (YB) areas in the Sichuan Basin develop a huge number of platform margin reef complexes, which have emerged as an important field for oil and gas exploration . The first study case is the Permian Changhsingian formation of the LG 3-D seismic data (991 [inline] × 1187 [crossline] × 501 [time]) shown in Figs. b and a. We employ the prediction workflow (Fig. a) with a sliding window traversing all the 3-D target strata, yielding the seismic facies classification result shown in Fig. c. In addition, we display the predicted results with different 2-D profiles in the crossline and inline directions in Fig. b–i. The regions indicated by the blue arrows are correctly predicted to be hummocky facies, which is roughly consistent with the geological structural uplift in the corresponding 2-D seismic profiles. However, some artifacts or inaccurate predictions still appear in some areas indicated by the red arrows in Fig. f and h, which is mainly due to the incomplete flattening of the strata. As shown in Fig. i, we can clearly observe a distinct reef–top interface reflection axis indicated by blue arrows. The trend of this reflection axis indicates that the reef gradually moves backwards in the increasing crossline direction, which closely matches the trends observed in both our predicted results and the expert interpretation results. Finally, we obtain the corresponding sedimentary facies result (Fig. d) based on the predicted seismic facies result, well log information, seismic data, and geological and geophysical knowledge. Our final sedimentary facies result (Fig. d) is highly consistent with the expert interpretation of sedimentary facies shown in Fig. e.

The second study case is the Permian Changhsingian formation of the YB 3-D seismic data (1300 [inline] × 475 [crossline] × 600 [time]), as shown in Figs. f and a. The YB data consist of more complex geological structures compared to the LG 3-D seismic data. Using the same prediction workflow as the previous case, we obtain the corresponding distributions of seismic facies and overlay the result with a manually interpreted horizon shown in Figs. g and b. The predicted distribution of hummocky seismic facies is consistent with the uplifted areas on the manually interpreted horizon. This high consistency can also be demonstrated in Fig. b–i, which display additional 2-D seismic profiles with the predicted result in a different 3-D view. In particular, the areas indicated by the blue arrows in Fig. g and h demonstrate that our method accurately captures the distribution of the patch reef and platform reef zone. Moreover, our predicted results are consistent with expert interpretation results of the reef distribution along the complex platform margin. However, some regions indicated by the red arrows (Fig. f) are incorrectly classified as other seismic facies, which is probably due to the unsuitable scale of the sliding window for these local regions, the influence of boundary effects from the sliding window, and incomplete flattening of the stratum. Finally, we also generate the corresponding sedimentary facies results shown in Fig. h, where the platform margin reef is clearly and reasonably resolved and the spatial distribution of the platform margin is highly consistent with the expert interpretation in Fig. i.