Generalization of deep learning models for natural gas indication in 2D seismic data

Methods based on Machine Learning and Deep Learning are increasingly popular to help interpret large volumes of data that belong to various areas and seek to fulfill multiple tasks. One of these areas stud-ies seismic data in the search for hydrocarbon reserves, for which Deep Learning models are trained, showing acceptable results for low study data. However, these models present generalization problems. Their performance tends to decrease when used on seismic data from new exploration. This tendency is particularly true for 2D data, which have many features. This work presents a method to improve the generalization of the Deep Learning model for the indication of natural gas in 2D seismic data based on the recommendation of training data and hyperparameter operations of the model. The tests used a database of the Parnaiba basin in northeast Brazil. Experiments showed an increase in the correct in-dication of natural gas that varies according to the metric 8% <= Recall <= 37% , with a fluctuation in the increase of false positives of -2% <= Precision <= 13% . It is an improvement in the generalization of the Deep Learning model of up to 11% according to the F1 score metric or up to 10% according to the IoU metric.(c) 2023 Elsevier Ltd. All rights reserved.

Automated summary

Methods based on Machine Learning and Deep Learning are popular for interpreting large volumes of data across various areas and tasks. This work proposes a method to improve the generalization of a Deep Learning model for indicating natural gas in 2D seismic data by recommending training data and hyperparameter operations. The experiments show an increase in the correct indication of natural gas by 8% to 37% and a fluctuation in the increase of false positives by -2% to 13%, resulting in an improvement in the generalization of the Deep Learning model of up to 11% according to the F1 score metric or up to 10% according to the IoU metric.