Banca de QUALIFICAÇÃO: GALLILEU GENESIS PEREIRA DE SOUSA

Uma banca de QUALIFICAÇÃO de DOUTORADO foi cadastrada pelo programa.
STUDENT : GALLILEU GENESIS PEREIRA DE SOUSA
DATE: 27/07/2022
LOCAL: Pos Graduação Engenharia Civil
TITLE:

Machine Learning-Based Models Applied to Case Studies in Complex Scenarios for Petrofacies Classification and Pore Pressure Prediction

KEY WORDS:

Reservoir modeling, machine learning, petrofacies classification, pore pressure prediction

PAGES: 200
BIG AREA: Engenharias
AREA: Engenharia Civil
SUMMARY:

Pore pressure estimation, which is controlled by lithofacies variation and diagenetic evolution, represents a critical issue for operations related to exploration and development of oil and gas reservoirs. These activities include drilling and completion of wells, and the identification and characterization of reservoir intervals. These themes present several challenges due to the lateral variation of the geology and the variation in the volume and distribution of data collected from wells. The search for more reliable estimation techniques has led to a growing interest in using Machine Learning algorithms to obtain pore pressure models and classify lithological patterns (petrofacies, lithofacies). For pore pressure prediction, most applications found in the literature are based on Deep learning models, which involve complex implementation, deficiencies in the interpretability of results, and high programming time and computational costs; in addition, these models do not present a competitive advantage for structured data problems, especially compared to Boosting ensemble algorithms. Classification of lithofacies proves particularly challenging in thin reservoir intervals with limited sampling or a reduced number of wells. This work used data from outcrops and wells to deal with complex aspects such as the thickness of the sampled interval or the regional distribution of wells, and to verify the efficiency of models based on artificial intelligence, classical models and models that have not yet been used for these problems. We have used the GNB and SVM techniques for the task of classifying petrofacies in a thin interval of laminated limestones with a high frequency of vertical variation. The latter obtained the best results in the classification of the main facies patterns (mean f1-score of 0.47). However, the GNB model presented the best performance when the analysis was focused on identifying the two main groups of petrofacies. We have also compared the efficiency between ANN models and boosting algorithms for pore pressure prediction on a regional scale based on geophysical logs of a few wells. In addition to GBM and XGBoost, we have included LightGBM, CatBoost , HistGB, and NGBoost models, whose application has not yet been described in the literature for this type of problem. Although we observed no significant differences regarding the scores of the models (except the GBM), the ANN and HistGM models obtained the most promising results, with the lowest generalization error. However, the ANN showed the worst results regarding scalability, with a higher computational cost, compared to Boosting models. The analysis showed that for these problems commonly treated by the oil industry, boosting-based approaches can provide faster development of routines, and lower computational cost with equivalent or better efficiency to what can be obtained with ANN. 

 

 

BANKING MEMBERS:
Externo à Instituição - MOISÉS DANTAS DOS SANTOS - UFPB
Externo à Instituição - YOE ALAIN REYES PEREZ - UFRN
Presidente - 1218780 - PAULO ROBERTO MACIEL LYRA