Machine learning approach to hydrocarbon zone prediction from seismic attributes over “GEM” field, Niger-Delta, Nigeria

A computer programme (in Python language) was developed for the generation and performance assessment of predictive models, capable of combining relevant seismic attributes for reliable hydrocarbon zone prediction ahead of drilling. It attempts to resolve the problem of making accurate and efficient interpretations from a large database of derived seismic attributes. The research utilized post-stack 3D seismic volume for the delineation of structures and the generation of seismic attributes. Six faults were identified across the mapped horizons. Five seismic attributes were generated and exported from 3D seismic data as numerical values for machine learning analysis. The binary cross-entropy classification metric was used to evaluate the performance of the developed predictive models while an individual seismic attribute (Maximum Amplitude and Extract value) map was used to validate the predictive models. A correlation of well depth-to-top of selected horizons with the seismic depth slice surface was used for further model validation. The Multi-Layer Perceptron (MLP) model results enhanced the visibility of the other five hydrocarbon prediction zones. The MLP predictive model map gave higher precision of the predicted hydrocarbon zones over the Self-Organising Map (SOM) predictive model, thus reinforcing the confidence level of the former.