Evaluating Machine Learning performance for seismic-based porosity prediction using synthetic models

In Geosciences the interpolation of well-data to a multi-dimensional model is a complicated and subjective process. Lately, machine learning (ML) methods are considered for this task. This presentation investigates the effectiveness of ML methods for porosity prediction using the inversion of seismic data from synthetic geological models.

Well data from the F3-block, Dutch North Sea, are used for calculating three synthetic models: homogeneous and heterogeneous wedge, and a fault model. The ML methods used are Lasso-regression, Random Forest, k-nearest neighbors (KNN), and Neural Networks. Other parameters such as variation of well locations are also tested.

The 2D predictions show that well-locations are important for determining porosity in the homogeneous wedge. The use of geologic timelines improve the prediction significantly when noise is added to the synthetic seismic section. Random Forest and KNN perform well in the homogeneous wedge, Neural network works well in the heterogeneous wedge, and Random Forest narrowly outperforms the other methods in the fault model.

This methodology allows experimenting with different ML methods and parameters using synthetic models. The effect of the ML method and parameters is apparent both visually and numerically. It provides insight into which combinations of ML methods and parameters are most effective for predicting porosity in synthetic models and actual cases.