This project is a part of Midnight’s APGCE Geohackathon Project.

The good seismic interpretation can be derived from the good quality of seismic data. Therefore, the missing seismic data including wipe out zone due to the shallow gas can limits our understanding in subsurface structures. Furthermore, the missing data can be affected to the processing techniques which lead into the misinterpretation.

Currently with the growth of computing power, one of computer technology called Artificial Intelligence (AI) has drawn much interest to solve the geoscience problems. In missing trace prediction, one of AI branch called deep learning has shown tremendous result in prediction missing trace in seismic data compare with the traditional interpolation technique. Hence, this research present new technique called pix2pix to address this issue. Pix2pix algorithm using Generative Adversarial Network (GAN) as the backbone of the algorithm that enables the model not only to generate the prediction but also discriminate the similarity between the prediction and the real data. The figure below shows the detailed workflow of the method:

By using 1500 epochs or 25 minutes of training pix2pix model can accurately predict the missing trace on the seismic data with higher than 0.8 cross-correlation coefficient and highly preserved the amplitude spectrum. Furthermore, the prediction results still preserved the fault delineation and excluded the artifact effect that occurred on the original data.

Skills: Python, Numpy, Pandas, scikit-learn, Keras, Scipy