Aplicação de inteligência artificial e machine learning em dados litoestratigráficos e geofísicos das expedições do Programa Internacional de Descobertas Oceânicas (IODP)

Abstract

Specific computational tools help the geologist to identify lithologies and stratigraphic stacking in well drilling, reducing operational costs and managing the professionals practical work time, directing them to efficient data interpretations or even in the improvement of scientific research in the region. in geologically distinct regions. In this study, the application of machine learning algorithms for the supervised classification of lithologies was evaluated. Data from records of multivariate parameters in offshore wells were used, related to the International Ocean Discovery Program (IODP) with supervised and unsupervised data (images), with the creation of a context of hybrid application of algorithms, divided into two manuscripts. In manuscript I, through the analysis of the lithologies proposed in 7 IODP-Expeditions and the use of the use of the algorithms, it was possible to group and divide the lithological sets into four groups of lithologies and templates. The geophysical properties used in the present study included GRA, PWL, MS, RSC and SRM. The templates were submitted to training and testing by the Multi-Layer Perceptron (MLP), DecisionTree, RandomForest and Support Vector Machine (SVM) methods, using the classification metrics as the result evaluation. As a result, it was observed that Template1 obtained better results in the MLP algorithm, Template2 and Template3 obtained better results in the RandomForest algorithm above 80.00% accuracy. For Cross-validation, the RandomForest algorithm achieved excellent performance in all scenarios. In the Practical Template, the G2 lithology group obtained the best result with the MLP algorithm with an average accuracy greater than 85.00%. For manuscript II, the division of the data included the formatting of three datasets: dataset0, dataset1 and dataset2, specifically with the data obtained during the IODP-Expedition 362. The petrophysical data used included PWL, GRA, RSC, NGR, MAD, MS, RGB and high-definition images. Dataset0 included temporary training data to validate the best interpolator. The dataset1 has the interpolated data of the petrophysical properties, making a total of 295,945 records for U1480 and U1481 with 17 features. The dataset2 covers the texture and color data extracted from the segmentation of the images, making a total of 85,058 records for the U1480 and U1481, with 90 features. Each data set is replicated in two groups of lithology: Group 1 and Group 2. For dataset2, new combinations are added between the features forming 102 practical arrangements with specific results in each combination. The values were interpolated by Linear, Spline, Slinear, Quadratic, Cubic, Akima, Pchip and Piecewise. The machine learning method used for all datasets is RamdomForest. The results show that the best interpolator evaluated in dataset0 is Akima with an accuracy of 98.22%. For dataset1, U1480, the accuracy value is 96.96% in the combination of 70% training and 30% testing in Group 1 and 97.71% in the combination of 70% training and 30% testing in Group 2. For dataset1, U1481, the accuracy value is 99.68% in the combination of 80% training and 20% testing in Group 1 and 99.74% in the combination of 80% training and 20% testing in Group 2. For dataset2, the best evaluated practical arrangements are 51 (Group 1) and 102 (Group 2) for U1480 and practical arrangements 32, 33, 44, 45 and 51 (Group 1) and 102 (Group 2) for U1481. Regarding the new Area Superpixel (Apx) method, the best results are in Group 1 with the greatest combination of training and the least combination of tests. The evaluated datasets were grouped in an organization between the expedition sites, in which a context of real practical application in the daily activity of the geologist was generated with excellent results of lithological classification. The interpolation of petrophysical data is valid and necessary when there is little data for training respecting the characteristics of each property and interpolator. Properties extracted from the images are relevant and grouped together with the petrophysical properties create a context of extreme importance in the discovery and presentation of information to the geologist.