Sensitivity analysis of similarity learning models for well-intervals based on logging data

Abstract

The problem of the interpretability of neural network predictions is crucial for industry, especially in areas where the cost of a mistake is incredibly high, to which oil&gas belongs. In this sphere, similarity models play an important role. One of the pre-required input data for constructing a geological model of hydrocarbon field are interwell correlation results through well logging data analysis. Recent publications address the problem of similarity assessment between distinct wells, which is the manual, time-consuming, and expert-based process. To enhance the interpretability of the developed deep learning models, we propose a visualization tool consisting of two methods. Our first method is an adaptation of a saliency map, which has already shown its visualization and interpretability quality. However, as the possibility of dealing with black-box models exists and is high enough, we propose another method based on the masking of the original well-interval. The case study from the Norway basin and our tool’s visualization quality evaluation demonstrate the effectiveness of the developed tool.