Full text
| |
| Author(s): |
Bruno Wamzer Jeiss
Total Authors: 1
|
| Document type: | Doctoral Thesis |
| Press: | Campinas, SP. |
| Institution: | Universidade Estadual de Campinas (UNICAMP). Faculdade de Engenharia Mecânica |
| Defense date: | 2022-11-17 |
| Examining board members: |
Alexandre Campane Vidal;
Michelle Chaves Kuroda;
André Pires Negrão;
Giovanna da Fraga Carneiro;
Aline Maria Poças Belila
|
| Advisor: | Alexandre Campane Vidal |
| Abstract | |
To enable petroleum exploration and production in the complex reservoirs from the Brazilian Pre-Salt, the development of consistent geological and petrophysical models are essential. The main goal of this research is to develop reservoir characterization workflows, by combining traditional geology and petrophysics, with machine learning. The studied interval in the Barra Velha Formation displays a succession of in-situ facies, consisting of shrubstones, laminated mudstones and spherulitestones, and reworked facies comprising intraclastic grainstones and rudstones. The first part of this study integrates core, thin sections and Acoustic Borehole Image Logs to segment the studied interval in Reservoir Petrofacies. Subsequently, reservoir quality is evaluated with nuclear magnetic resonance (NMR) and Micro-CT. The best reservoir quality belongs to rudstones with low cementation, and grainstones with moderate dolomite/quartz cementation, with values for porosity between 15-20% and permeabilities between 100 and 500 mD. The best in-situ facies belong to shrubstones with moderate dolomite cement, with porosity between 12-17%, and permeability around 100 mD. This integrative framework has shown to be effective in transferring reservoir quality information from core to well log scale, allowing the interpreter to make inferences outside the cored interval, respecting geological constraints. The second part of the study addresses imprecise porosity estimation from well logging tools, especially in sections dominated by secondary porosity. This study applies supervised machine learning, integrating data from conventional logs, NMR and a vuggy index extracted from Acoustic Borehole Image Logs. The results are verified against routine core petrophysical analyses. Random Forest, XGBoost and Support Vector Regression algorithms were tested. The performance metrics from the three models are R2 of 0.835, 0.836, 0.800 and root mean squared error of 1.75, 1.68, 1.82, respectively. All three models outperformed the original measurement from NMR with R2 of 0.64 and RMSE 2.78. In addition to predictive performance, the method offers cost and time-saving advantages, when compared to conventional methods. Methodological hurdles may arise especially due to artefacts in acoustic image logs, which need to be carefully mitigated prior to model training (AU) | |
| FAPESP's process: | 20/01320-2 - Petrophysical analysis of a pre-salt carbonate reservoir using well logs |
| Grantee: | Bruno Wamzer Jeiss |
| Support Opportunities: | Scholarships in Brazil - Doctorate |