3.1. Collect large corpora of representative training data

We had access to several seismic data volumes (each survey spanned across hundreds of kilometers) and salt interpretations, acquired over long periods of time (legacy data) and from many different geographical areas. These are used to create a large corpus of training pairs consisting of seismic and labels derived from existing salt interpretations. To accommodate large variability in acquisition and processing techniques, the training data were selected across these variables so that the trained models are robust to these variabilities.

Subsurface geology and salt geometries vary depending on geographical regions. In addition, salt is interpreted differently depending on the underlying local geology, so labels used for different regions are also distinct, in terms of their alignment with the underlying seismic signal. Hence, different DL models were trained for different geographical regions using training data from the same geographical regions. For instance, training a GoM specific DL model was trained using data from only GoM assets encompassing different geologies and subsurface geometries that exist within GoM. This GoM model was then deployed for seismic interpretation throughout all the assets across GoM.

3.2. Create training data from 3D seismic volumes

Seismic data typically has quite a large dynamic range of amplitudes. The amplitude distribution resembles a normal distribution with long tails, as shown in Figure 3 (image on the right). The salt edges coincide with either positive (shallower salt edges) or negative (deeper salt edges) peaks in amplitudes. In addition, depending on the type of acquisition and processing, the datasets go through the dynamic ranges across datasets vary widely as well. Numerous data normalization techniques were implemented, including log scaling, z-score scaling, and scaling to range (−1 to +1). Empirically, we found that simple scaling to range (−1 to +1) worked best, as it preserved the complete dynamic range of each dataset which was essential to retain the discrimination between dominant edges and other edges in the data.

Figure 3. 2D seismic slice from a 3D seismic dataset and the distribution of the seismic amplitudes (wide dynamic range).

The goal of a salt interpretation DL model was to learn the mapping between seismic and salt edge. Seismic data are typically a 3D volume with the x and y axes representing the lateral cross sections and the z-axis representing the depth. In this effort, to align the DL models with top-down salt interpretation workflow, a predominantly 2D strategy was adopted, where 2D “tiles” were extracted from the lateral slices. Depth slices were not included in the training data. A large dataset of 2D tiles of predetermined size was extracted from cross sections of seismic volumes and the corresponding labeled volume.

3.3. Training generalized salt interpretation DL models

Salt interpretation models were designed to identify and delineate the salt-sediment boundary effectively. In the initial training phase, we started with fully connected CNNs and U-nets (Ronneberger et al., Reference Ronneberger2015) and by incrementally adding datasets to the training set. The first goal of the model was to boost salt edge detection with high accuracy. High accuracy has two components:

1. 1. Detect salt with high overlap with the labels, and

2. 2. The position of the detected edge consistently aligns with the seismic loops.

For top-salt and TSOH models, the salt edge should consistently align with the positive peaks in the seismic signal, whereas for base-salt and BSOH models the salt edge should consistently align with the negative peaks in the seismic signal.

Given these established criteria for the DL models, we incrementally started to incorporate more data from more diverse surveys. Training the model with more data helped in boosting the salt edge detection along areas with strong seismic signature; furthermore, adding more diverse data also improved the generalization ability of the model. However, we also noticed two degradations in the quality of detections:

1. 1. Consistent increase in the persistent false salt detection (or false positives) and they increased with the amount of data added to training. An example of the false detections is highlighted in the yellow ovals in Figure 4, panel b.

2. 2. Incorrect positioning of some of the salt detection, especially in areas of complex geology and areas of low SNR in the seismic, like overhangs. An example of the incorrectly positioned detections is highlighted in the black circle in Figure 4, panel b, as compared to the salt label in panel a.

Figure 4. The figure shows a comparison of multiple top-salt models (c and d) re-trained using model-centric and data-centric methodologies to improve top-salt (b) detection inaccuracies.

To address the false salt detection, the number of 2D training tiles where no salt exists, hence the no label, were added to the training data used to train the models. In general, the number of non-salt used was about a forth the total number of tiles with salt. This provided a reasonable balance between 2D tiles of positive and negative examples of salt detection for the model. This strategy reduced the deeper false salt detections significantly, yet the shallower ones persisted.

To address the complexity in the geology, we introduced more sophisticated neural network architectures like, fusion net (Quan et al., Reference Quan2016), dense net (Huang et al., Reference Huang2017), and pyramid scene parsing (PSP) networks (Zhao et al., Reference Zhao2017). Simply going from U-net to PSP architecture does not allow to train a model that addresses the in-correct salt detection, as is evident from panels b and c of Figure 5, highlighted by the black circle. However, going to a more complex architecture did remove the shallow false detections, at the same time resulted in breaks in the detected edge as well.

Figure 5. The top image shows the overlay of the thick top-edge label used for training, and the bottom image shows the same label to highlight the top-edge of label is aligned with seismic peaks.

As mentioned in the introduction section, another requirement of the DL models in terms of model performance is that they should be integrated into the existing salt interpretation workflow. This requirement renders model’s performance characterized by traditional metrics like accuracy is less meaningful. Rather the model’s performance at creating an inference best suited to maximize the performance of any downstream tasks in the workflow is most crucial. In the context, the inference generated by the trained DL model needs to be converted into a 3D surface data structure. Hence, for the process of conversion of inference to 3D surface to be effective, it is more important that the DL model produce as little false detections while being highly accurate in coverage and positioning of the detected edge.

Even with increasing the amount and diversity of training data and adopting more complex neural network architectures, the false salt detections in incorrect positioning of detected salt persisted making the models well-generalized yet unsuitable for production.

4.1. Label quality

The salt interpretations that are used as labels are typically single-valued edges along seismic peaks. We found that these labels were more susceptible to errors in the labeling process. The labels were hence made thicker, as shown in Figure 5, to alleviate incorrect or imprecise labels. We experimented with several label thickness values and determined that any thickness value that reliably encompasses one to two adjacent seismic loops boosts robustness of prediction positioning and better localization on the loop of interest. This also acts as implicit regularization and helps generate wide range of prediction probabilities, mainly in areas of ambiguity in the seismic edges.

4.2. Seismic quality along salt edge

We observed that there were two areas where the DL model struggles, either to detect salt resulting in loss of salt coverage or incorrectly positions salt.

1. 1. Areas of complex geology, which manifests as ambiguous seismic signature, as highlighted by black circle in panel b in Figure 4.

2. 2. Areas of weak or no seismic signature, as highlighted by the red circles in panels c and d in Figure 6.

Figure 6. Figure shows the seismic panel with different salt edge labels overlaid, images c and d showcase the salt overhang labels and the complete salt boundary labels to highlight the structural context which enables DL models to learn effectively.

In the model-centric approach described in Section 3, for a given 3D seismic volume, all the tiles created from that volume where salt label exists, in other words, salt exists, are used for training. In response to the above observations, we adopted a slightly different strategy during the selection of training tiles, tiles where geology was sufficiently complex or seismic signature was weak/non-existent but contained salt labels were deliberately excluded from the training.

As the percentage of data with weak or ambiguous seismic signature for salt labels reduces, the position accuracy of trained model in complex geology or ambiguous seismic improved. This improvement in positioning is illustrated within black circle in panel d of Figure 4. Within the black circle, neither the U-net (panel b) or PSP-net (panel c) produces a detection aligned with the label (panel b). With data selection based on seismic quality and increasing thickness of labels, even with comparatively less complex architecture like U-net the salt edge positioning is accurately placed.

In addition, in other complex areas, the model simply did not generate detections instead of generating erroneous ones. Thus, removing data where the labels were not supported with clear seismic signatures from training data enables the model to learn the salt-sediment boundary effectively, as the model is more robust to ambiguous seismic as it relates to interpretations (or labels).

To operationalize this strategy in the training data generation pipeline, we filter out tiles where the labels are significantly inconsistent with the seismic loops or entropy of the seismic data underneath the label is below an acceptable threshold. This also generated models that are robust to small perturbation in seismic alignment and the presence of noise, thus data with smaller alignment deviations could still be used to reliably train. This allowed us to leverage quite significant sections of large seismic surveys into training. In addition, we pruned out the tiles that only have a smaller salt percentage within the tile.

4.3. Sufficient structural context

However, the detection of salt where seismic signature is weak or non-existent persisted. For models trained to detect steeply dipping edges of the salt body (called overhangs) which are not very common and tend to be poorly imaged, very few representative datasets with strong seismic signature were available.

Even though the salt overhang labels were not entirely supported with well-imaged seismic, if we leverage labels that outline the entire salt boundary, which we refer to as salt rings, within a tile, there will be enough area of the label that has a strong seismic support. Such labels also provide a structural context to the salt boundary. For the overhang specific DL models, the labels were extended to include not just the steeply dipping edges but the entire salt boundary. Panels c and d from Figure 6 show a visual comparison of steeply dipping salt edge labels with structurally more complete salt boundary labels.

Just a few datasets, with enhanced salt structure context labels along with standard data augmentation techniques, was a significant breakthrough in effective model training. Figure 7 Illustrates the enhancement in the DL models’ performance after training with entire salt ring labels. The yellow arrows in panels a and c point to the areas of the overhang that the original DL model was unable to detect the salt edge; however, if the DL model is trained with sat ring, which embeds the salt’s structural context, the overhang coverage is boosted significantly. Better coverage on the overhang makes the downstream task of 3D surface extraction also simple to complete with high accuracy.

Figure 7. Figure shows two pairs of seismic panels with salt overhang detection coverage (panels a and b) of DL model trained using only overhang labels compared to model (panels c and d) trained using salt ring labels. It showcases the salt overhang coverage is boosted significantly using salt ring labels to train the models.

Thus, enhancing the quality of the training data by systematically focusing on training data generation that boosts label context, reduces label ambiguity, and increases seismic support for the label improved the model performance. The focus on optimization of the training data and label quality to improve the DL model’s performance makes these strategies data-centric rather than model centric and enables the well-generalized model to be deployed in production and integrated into the salt interpretation workflow.

The DL models, which turned out to be the most robust and suitable for deployment, were trained with much less training data still encompassing the geological variability, however, with much higher quality. In this context, high quality refers to the data that allow the models to learn effectively; in other words, data that allow the model to learn robust representations of the task it is learning. For salt interpretation, these observations were incorporated into automated training data generation through systematic label generation and data quality filters.

5.1. Data quality criterion for model performance prediction

Most of the salt interpretation DL models were trained on legacy datasets and during deployment, applied to data from new terrains which are acquired and processed through the most recent acquisition and processing techniques. This further adds to widening data drift the salt interpretation models encounter.

Given the data-centric strategies applied to generate the training data, it is reasonable to expect that data where the quality of seismic image at salt boundary is strong, the DL models will perform well, in other words, high coverage and accurate positioning of detected salt. In addition, salt interpretation was posed as an edge detection problem, so the DL models were optimized to detect the salt edges. Hence, it follows that the model will be better able to delineate the salt edge where the edge contrast is high. Based on this reasoning, we heuristically designed a set of criteria that quantify the minimum data quality required to perform the model’s task manually.

This criteria, called data conditioning criteria, consists of two parts: one is the lower bound on seismic data contrast (quantified using RMS contrast) and second, detects seismic loop clipping. The data conditioning criteria has served in practice, as a reliable predictor of a trained model’s performance on a blind dataset, as it roughly quantifies whether the test data sufficiently capture phenomena that model needs to generate acceptable performance. In addition, the criteria also provides the interpreter, with guidance to improve the blind/test data quality through processing techniques that will allow the salt interpretation DL model to better delineate the salt-sediment boundary, in effect, providing a data quality sufficiency test.

The ability to predict the DL model’s performance and guidance to improve data quality both contribute significantly to boost the interpretability of the DL model.

Figure 8 shows the evolution in the seismic contrast and seismic loop clipping, as the seismic is processed with different filters. As is evident as the seismic processing diminishes the seismic signature along salt edges, we see the RMS values evolve and detected salt coverage is also affected adversely.

Figure 8. The panels from left to right showcase the original seismic data, de-convolved seismic and Laplacian filtered seismic. The plots on top right of each panel shows the evolution of the data-conditioning criterion associated with each seismic.

Furthermore, the data conditioning criteria allow to loosely detect a domain gap as the model is deployed in real world. Domain gap refers to the large difference between the model training data and the model test data and is quite difficult to understand. Thus, the metric can potentially be leveraged to automate the selection of high-quality data for continuous improvement of models in production. In effect, providing a lower bound for training appropriateness analysis of new datasets and enabling re-training of DL models to boost longevity.

5.2. Statistical data quality testing to predict model failure modes

The data-centric training of DL models has underscored the importance of seismic data quality in developing reliable, well-generalized models. Hence, it follows that DL model’s salt detection performance is also highly correlated with seismic quality, particularly as it related to salt edges or salt-sediment boundaries. As seismic quality diminishes, the model performance will also degrade, and to better understand the limits of the DL model performance, we implemented statistical tests, which aim to quantify an “operating range” of seismic quality beyond which the model’s seized to detect salt. This in effect provides the failure cases for the DL models.

The statistical tests are designed, given an indication of how robust the DL model performance is to statistical variations in test data. For these statistical tests, a set of blind datasets that vary from the training data in terms of acquisition and processing techniques were selected as the test corpus. Subsequently, these datasets were subjected to a series of common seismic processing algorithms to simulate effects real-world processing on quality of seismic imaging. Figure 8 shows the 2D slices of the original seismic data and the corresponding processed slices with two different processing techniques. This provided a set of data that represent the variations in seismic quality that the models may encounter in real-world settings. The salt interpretation model’s performance over these seismic variations was used to empirically determine a set of processing algorithms and the associated parameters which diminished the model’s performance beyond the acceptable range. These tests further boost the DL model’s interpretability by determining an “operating range” where the model will provide reasonably accurate salt detection.