Dongfang Qu, Klaus Mosegaard, Runhai Feng, Lars Nielsen

Numerous studies have demonstrated the capability of supervised deep learning techniques for predicting geological features of interest from seismic sections, including features that are difficult to identify using traditional interpretation methods. However, successful application of these techniques in practice has been limited by the difficulty of obtaining large training dataset where seismic data and corresponding ground truth labels are well defined. Manually creating large amounts of labels requires a heavy workload, and the uncertainty of the interpretation and labeling process decreases the model’s ability for making accurate predictions. Using the chalk-flint sequence scenario onshore Denmark as an example, we present a novel workflow of generating large quantities of synthetic training data with high-quality labels using stochastic geological modelling, and we investigate the capability of a synthetic data-trained convolutional neural network for predicting sub-resolution thin layers from seismic sections. It is shown that a neural network trained on synthetic data can predict a realistic number of sub-resolution flint layers from real seismic data that have been collected from the Stevns region in Denmark, which has value for understanding of overall geological characteristics of the succession and engineering applications such as construction site evaluation.