Lithology classification based on well log data has attracted the attention of geologists, as the extraction of rock samples from the wellbore is a costly process. Although some machine learning techniques have recently been applied to the problem, the existing studies are not standardized in terms of data processing and evaluation protocols. This paper aims to bridge this gap by providing common ground for the comparison of prospective works on lithology classification from well logs. The proposed benchmark is inspired by a thorough review of the literature, from which the main aspects of an experimental protocol are captured under the principles of simplicity, variety, impartiality, and meaningfulness. The public datasets used in the experiments are the FORCE and Geolink repositories. Ten methods are tested, including six deep learning models and four shallow machine learning algorithms, and compared by nine distinct performance metrics under cross-validation. Different sequence lengths are tested as input for deep learning models to assess their contextual understanding capabilities. The experimental results using the proposed setup show that context-insensitive models outperform most context-sensitive ones, but with high correlation. The study also reveals that the problem of lithology classification is still not solved. Therefore, the benchmark presented herein serves as a fair protocol to assess different methods and can be used by future works as a simple and unbiased way to compare new approaches in lithology classification.

The capacity to predict the occurrence and quality of source rocks in a sedimentary basin is of great economic importance in the evaluation of conventional and non-conventional petroleum resources. Direct laboratory examinations of rock samples are the most accurate way to obtain their geochemical properties. However, rock information is usually sparse, and source rocks are often sampled at positions that may not be representative of the average organic content and quality of oil kitchens. This work proposes a work flow supported by machine learning methods (random forest, DBSCAN, and NGBoost) to automate the source rock characterization process to maximize the use of available data, expand data information, and reduce data analysis time. From the automated quality control of the input data through the extrapolation of laboratory measurements to continuous well logs of geochemical properties, culminating in the 3D estimation of these properties, we generate volumes of total organic carbon (TOC) by applying machine learning techniques. The proposed method provides more accurate predictions, reducing uncertainties in the characterization of source rocks and assisting in exploratory decision making. This methodology was applied in the presalt source rocks from Santos Basin (Brazil) and allowed us to quantify the TOC distribution, improving the interpretation of the main source rock interval top and base based only on seismic amplitude data. The result suggests higher TOC values in the northern and western grabens of the studied area and a higher charge risk in the eastern area.

Traditional approaches for seismic characterization of source rocks use simple regressions between acoustic impedance and TOC (Total Organic Carbon) to calculate volumes of this property. In this pioneering work, we generate volumes of TOC and hydrogen index (HI), applying machine learning techniques, from elastic inversion outputs. By using multiple attributes and allowing the fitting of nonlinear models to the data, machine learning techniques provide more accurate predictions of these properties, reducing uncertainties in the characterization of source rocks. The methodology proposed in this work uses machine learning methods to automate all the source rock characterization process steps: from the automated quality control of the input data; through the extrapolation of sparse laboratory measurements to continuous well logs of geochemical properties; and culminating in the 3D estimation of these properties.