Abstract
This paper discusses a methodology to enhance machine learning (ML) models developed to predict lithology from real-time mud logging data. Inaccuracies of geological data obtained in the field can lead to inconsistencies in model predictions. To address this, drilling cuttings collected from the field were transported, re-described, and re-evaluated by a team of geologists in a laboratory environment with more favorable conditions. The analysis process carried out was composed of consistent and robust analysis methods and processes, such as microscopic examination (for rock type identification, percentage of each lithotype, color, texture, hardness, cement/matrix material, accessory minerals and fossils, sedimentary structures, and visual estimation of porosity), HCl (10%) test, and fluorescence to describe the cuttings samples. Additionally, calcimetry was used to determine the percentage of carbonate in the rock samples. The results of this study demonstrate that a more accurate reinterpreting of cuttings samples can increase the predictive capability of the ML model (which was able to estimate expected lithotypes with greater reliability). It is worth noting that the continuous description of new samples from other wells in various geological contexts of the same basin is fundamental to improving model performance and making it more trustworthy and representative of reality. One of the importance of these models is to assist the geologist in estimating the rock type being drilled in real-time, based on drilling parameters and gamma-ray data. In this sense, the model becomes more reliable during the decision-making process in drilling operations, contributing to the effectiveness, efficiency and automation of exploration for oil and gas fields. These findings have important implications for the development and implementation of ML solutions in the O&G industry, highlighting the importance of continually improving geological databases to enhance the accuracy and reliability of such models.
