Data-Centric, Interactive Deep Learning for Complex Geological Features

Data-Centric, Interactive Deep Learning for Complex Geological Features: A Groningen Case Study

Scotty Salamoff, Julian Chenin, Benjamin Lartigue, and Paul Endresen


Detailed interpretation of complex facies intervals within high-resolution 3D seismic data is a tedious and time-consuming process, even with the assistance of traditional deep learning methods. Traditional windowed waveform classification algorithms can have a non-unique solution and are impacted heavily by interpreter bias and laterally varying data quality. This is especially the case in deformed facies intervals such as post-depositional deformation of complex geological sequences, where tectonic reactivation and/or salt tectonism have re-worked sequences of post-salt siliciclastics into complicated packages that are difficult to interpret. These heavily re-worked zones subjected to erosion and re-deposition of growth sequences are prolific throughout the North Sea and can play an important role in fluid migration and containment. The sediments within the sequence of interest are post-depositional to original salt deposition and syn-depositional to halokinesis. Their complexity usually means such sequences are under-interpreted, which introduces pre-drill uncertainties about the well path or target itself (Tang et. al. 2012).

Therefore, we propose a new, data-centric, and interactive deep learning methodology using InteractivAI, which leverages neural networks to accurately yet quickly predict separate deformed facies in the Groningen study area. The results presented below were obtained in a fraction of the time compared to traditional interpretation workflows and allow geoscientists to better characterize complex geologic units while also determining potential impacts on prospective petroleum systems or planned well paths.


The Groningen study area is located onshore The Netherlands (Figure 1a). The gas reservoir area is characterized by a classic four-way closure with salt providing a competent regional top seal and lateral stratigraphic pinch-outs providing closure. Regional geometry is shaped by North Sea extensional/reactivation tectonics and local structure controlled by salt movement. The Upper Rotliegend Group was deposited during the Permian in a broad basin and is locally comprised of the entire sequence above the Base Permian Unconformity (BPU) and below the overlying Zechstein Formation (Figure 1b) (Boogaert 1976). As the trend is followed eastward, an older Lower Rotliegend Group sequence made up of volcaniclastic sediments is also present in extensional-fault-bounded grabens.

Figure 1: A) Location map of the Groningen High, The Netherlands (modified after Kortekaas et. al. 2017) whereas B) highlights key geologic intervals on an interpreted inline.

Available Data

The seismic dataset is a post-stack, depth migrated onshore survey that is SEG negative polarity, where the sea floor is observed as a peak. To best demonstrate the proposed methodology, the volume was cropped volume without vertical masking or additional sub-setting. Publicly available well data with complete log suites were used as the basis for initial labeling and were carried out from the well location as far as could be visually determined. The Siddeburen wells were selected for their robust log data and seismic ties.


Many recent studies (LeCun et al. 2015; Bandura et al. 2018; Chopra and Marfurt 2018; Chenin and Bedle 2022, Salamoff et al. 2022) have shown promising applications of machine learning techniques that aid in the recognition and classification of different seismic reflection patterns. In this study, we employ a data-centric, interactive deep learning approach, where data labeling/classification, network training, and class predictions happen simultaneously. This is only made possible by leveraging the VDS format, which can adaptively stream randomized images directly into TensorFlow.

Geoscientists provide active positive and negative reinforcement feedback to a binary or multiclass deep learning network during the training process and before generation of final classification probability results. Masking and validation methods were used to focus the training on relevant data and mitigate overfitting. Geologically-accurate and detailed salt and deformed facies models within the Groningen study area were generated by using a combination of interactive feedback reinforcement, data masking, and validation techniques.


A total of 7 inlines (ILs) and 6 crosslines (XLs) were labeled for salt identification (~0.38% of the dataset). The network type used was an interactive, binary U-Net with a Weighted Root Mean Square Error (RMSE) loss function applied. The model was trained for 14 epochs, where it took approximately 1.5 hours to train the network and generate the salt probability cube and automatically extract a 3D geobody (Figure 2a-b). This probability cube could then be used as a 3D mask to exclude from further detailed seismic deep learning (Figure 2c).