Predicting the magnitude of injection-induced earthquakes using machine learning techniques

Predicting the magnitude of induced earthquakes by underground injection is a critical strategy for risk assessment. This paper proposes the application of three machine learning techniques-support vector machine, probabilistic neural network, and AdaBoost algorithm-to predict the magnitude of the largest injection-induced earthquake (M) within a predetermined period. These machine learning techniques are used to model the relationships between ten input parameters-six seismicity indicators and four inputs related to injection wells-and earthquake magnitude classes (M < 3, 3 <= M < 4, and M >= 4). Models are applied to the earthquake and injection data for the Central Oklahoma region in the USA, and their input data are balanced using the data-level approach. The performance of each model is measured using the average recall of earthquake magnitude classes. The results show that balancing the training data improves the performance of the models, and the magnitude of induced earthquakes depends on the injection volume in the nine months before the earthquake prediction period. The parametric analysis of each model's input reveals that induced earthquake magnitudes are more likely to occur when there are shorter distances between the bottom of injection wells and the crystalline basement. Among the investigated models, the support vector machine model trained on the data balanced using synthetic minority oversampling technique performed best by predicting an average of 72% of earthquake magnitude classes. Overall, the findings of this study will allow for predicting the magnitude of induced earthquakes and the development of an early warning system for policymakers and residents living in areas prone to injection-induced earthquakes.

Automated summary

This paper proposes the application of three machine learning techniques to predict the magnitude of induced earthquakes caused by underground injection. The models are trained on earthquake and injection data from the Central Oklahoma region in the US, and their input data are balanced using a data-level approach. The results show that balancing the training data and considering the injection volume in the nine months before the earthquake prediction period improves the performance of the models. Among the investigated models, the support vector machine model trained on balanced data performs the best, predicting an average of 72% of earthquake magnitude classes.