Comparison of the effectiveness of machine-learning methods for solving the problem of quantitative prediction based on seismic data

Introduction. One of the key tasks for seismic interpretation is the prediction of the geological structure of the studied formations. In particular, a common task is to estimate the net thickness of reservoirs based on available well statistics. Such a task is standard in the framework of dynamic wave field analysis and is often solved by constructing a predictive model based on available geological and geophysical information, including values of net thickness in available wells.

Goal. The purpose of the work is to evaluate the effectiveness of machine learning methods in solving the problem of reservoir thickness prediction based on seismic data. Modern data analysis often uses this category of methods to build various predictive models. Seismic interpretation, in turn, is often associated with the use of relatively simple linear models. This makes it relevant to determine the gain from the use of complex prediction models.

Materials and methods. To carry out the study, a relatively well-studied area of one of the fields in Western Siberia was used. The territory under consideration is completely covered with 3D seismic data, there are 170 wells for constructing the model, in which the value of net thickness is determined.

To implement the study, both a standard linear regression and more complex machine learning algorithms are considered. Among the algorithms, multidimensional regression, random forest method, nearest neighbor method and neural network are considered. To assess the quality of prediction, the available sample of wells is divided into training and validation samples consisting of 80 and 90 wells, respectively. All calculations are implemented using open python programming language libraries.

Results. As a result, distributions of the expected accuracy of the forecast for each of the considered methods were obtained. The text of the article describes in detail the research algorithm, as well as the tests performed to select the parameters of each algorithm.

Conclusion. The results obtained allow us to conclude about the effectiveness of using machine-learning methods. All the approaches considered make it possible to obtain a more accurate prediction of the net thickness compared to the linear regression approach. The most significant increase in accuracy is observed with using a neural network and the improvement estimated as 23 %

1. Meckel L.D., Nath A.K. Geologic considerations for stratigraphic modelling and interpretation. In Seismic Stratigraphy — Applications to Hydrocarbon Exploration, ed. C. E. Payton. AAPG Memoir, 1977, no. 26, pp. 417-438.

2. Linear Models [Electronic Resource]. Access: https://scikit-learn.org/stable/modules/linear_model.html#lasso

3. Decision Trees [Electronic Resource]. Access: https://scikit-learn.org/stable/modules/tree.html#regression

4. Nearest Neighbors [Electronic Resource]. Access: https://scikit-learn.org/stable/modules/neighbors.html

5. Neural network models (supervised) [Electronic Resource]. Access: https://scikit-learn.org/stable/modules/neural_net-works_supervised.html