A investigation of data quality in reservoir characterizations using Machine Learning

Resumo

Inferring the capacity of reservoirs is one of the essential tasks in the oil and gas exploration process.
The characterizations of transport and storage are crucial for reservoir evaluations and, therefore, depend on the

permeability and porosity. Although estimating the permeability in porous media is challenging since experi-
mental data gathering is very costly, estimations are not accurate. Machine Learning (ML) methods have been

applied to predict the permeability in oil-producing areas as cost-effective and quick characterization strategies.

However, the quality predictions of ML algorithms depend on the available data quality and the algorithm param-
eters optimization. In this work, in order to have a comprehensive understanding, we investigate the permeability

inference employing algorithms as Multivariate Linear Regression, Decision Tree Regression, Support Vector Ma-
chines (SVM) and Multilayer Perceptron (MLP). The ML approach was constructed and tested via data samples

experimentally gathered from Australia and Papua New Guinea region. Data pre-processing metrics are optimized.
The most relevant feature was analyzed and optimized parameters improved the inferences as expected. The mean
squared error and root mean squared error for the test set are on the order of 0.0066 and 0.0811, respectively,
indicating that our results are very promising.