Petrophysical models have ruled the way to perform formation evaluation at well location for more than half century. In particular, open-hole logs represent the pillars for obtaining useful parameters necessary to optimize the perforation strategy, characterize the subsurface, and distribute rock properties in the 3D reservoir model. However, this conventional workflow could be highly problematic in reservoir revisions and data-room activities, when big amounts of data have to be handled in real-time, and often with limited information.
In this respect, oil companies have stored huge datasets for their assets, with quality-checked well log data and related petrophysical characterizations that intrinsically contain years of modeling and know-how in different scenarios. This paper deals with a novel data-driven approach aimed at obtaining a reliable formation evaluation by means of a fit-for-purpose machine learning algorithm. The latter is deemed able to automatically shed light on the statistical relationships hidden in the available log-petrophysics datasets, hence avoiding time-consuming processes and not efficient standard interpretations when the number of wells to be handled is large.
The implemented methodology takes advantage of the Multi-Resolution Graph-based Clustering (MRGC) approach that gathers its knowledge by recognizing patterns in well logs by means of non-parametric K-nearest-neighbor and graph data representation. This allows the system to learn through experience how log measurements are related to important petrophysical parameters (e.g. porosity, water saturation, and permeability). In detail, an automated screening can be performed aimed at detecting possible out-of-trend wells, and a homogeneous petrophysical revision carried out within a time-efficient template.
The potentialities of the MRGC method are demonstrated by a real case application consisting of data collected from more than one hundred wells drilled in a giant mature field. A subset of open-hole logs and petrophysical parameters coming from their modeling has been used as training set for the learning phase. Then, MRGC is used to predict the petrophysical responses for the logs belonging to a given validation set. In the end, a test dataset provides an unbiased evaluation of the comprehensive approach. The successful outcomes from the last step of the workflow show how, with a statistically representative and good quality dataset, machine learning can efficiently mimic high-skill expert work in harsh circumstances. Finally, the generalization to other relevant reservoir scenarios is also discussed.