The machine learning model achieves 97.97% accuracy in earthquake predictions in Los Angeles, California

Researchers in Georgia Southern University's Department of Information Technology in Los Angeles, California, have created a new foundation for earthquake prediction by developing advanced machine learning and neural network models built with a detailed feature matrix to maximize prediction accuracy .

They believe that applying a carefully developed feature set to the Random Forest machine learning model will provide accurate predictions of the maximum earthquake category in the next 30 days. Compared to the other 16 machine learning algorithms tested, Random Forest proved to be the most effective.

Professor Lei Chen, co-author of this study from the Department of Information Technology at Georgia Southern University, confirmed that this research opens new possibilities for the application of machine learning in disaster risk management and provides predictive tools that can help authorities prepare for risks and hazards .

“Integrating advanced machine learning algorithms such as random forest and neural networks has allowed us to break new ground in seismic prediction,” added study co-author Professor Yiming Ji, also from the Department of Information Technology at Georgia Southern University , added ,

“Our team's work not only pushes the boundaries of earthquake prediction, but also lays the foundation for future advances in applying machine learning to other natural disaster prediction models. The impact on improving public safety and emergency response is enormous,” said Professor Christopher Kadlec, associate professor at Georgia Southern University.

With an accuracy rate of 69.14% in predicting maximum earthquake magnitude within one of six categories, researchers in previous studies created a prediction pattern matrix for Los Angeles.

They expanded their research to Istanbul, Turkey, one of the most earthquake-prone zones, and achieved an accuracy rate of 91.65%. With further development, the accuracy rate for San Diego increased to 98.53%.

After successful results in San Diego and Istanbul, the researchers returned to Los Angeles to improve the accuracy rate from the previous 69.14%, this time reaching 97.97%.

They state that these results suggest that machine learning techniques could significantly improve the accuracy of earthquake prediction and provide authorities with a more effective way to prepare for risks and hazards.

“Our model’s 97.97% accuracy represents a significant improvement over traditional methods and provides important insights that can save lives and reduce property damage in high-risk areas,” said Cemil Emre Yavas, another co-author of the study.

The researchers used a combination of machine learning and neural network techniques to predict seismic activity in Los Angeles, relying on a dataset of earthquake reports from the past 12 years. Through advanced feature engineering, they were able to create a matrix that contains key predictive inputs.

The scientists used earthquake data from the Southern California Earthquake Data Center (SCEDC), which is managed by the California Institute of Technology.

Earthquake prediction accuracy can be improved by detecting deep seismic patterns, testing multiple tools, and studying seismic frequency features. Based on this study, scientists created and tested 16 different machine learning and neural network algorithms to select the most effective model for predicting the largest earthquake magnitude within 30 days.

This research aims to improve predictive modeling tools for the Los Angeles region by integrating insights from other earthquake prediction studies. Researchers are trying to increase the accuracy of earthquake prediction by combining machine learning algorithms, feature extraction methods and advanced neural network topologies.

A radius of 100 km (62 miles) was chosen to cover a wide area around Los Angeles, which is of great importance for earthquake prediction.

Using a radius of less than 100 km (62 miles) can exclude important seismic events that are critical to understanding earthquake patterns in the region.

Conversely, a radius greater than 100 km (62 miles) could introduce noise due to the inclusion of data from areas with different seismic properties and potentially reduce the model's prediction accuracy.

Thus, a 100 km (62 mile) radius represents an ideal balance that ensures sufficient data while maintaining the relevance and accuracy of the model.

When fine-tuned with appropriate hyperparameters, the Random Forest model provides robust and accurate predictions. These hyperparameters allow the model to exploit the full complexity of the data, thereby improving its predictive capabilities.

This fine-tuned model can support further analysis and could serve as a valuable tool in earthquake prediction.

The research builds on an extensive series of earthquake prediction studies conducted between 1990 and 2024.

Advanced neural network models, such as that of Bilal et al. Graph Convolutional Neural Network described, can significantly improve the performance of earthquake prediction. Their focus on early earthquake detection using complex neural network designs demonstrates the potential of advanced technologies to improve prediction capabilities.

Initiatives such as the Collaboratory for the Study of Earthquake Predictability (CSEP) and the Regional Earthquake Likelihood Models Experiment (RELM) by Schorlemmer et al. (2010) have paved the way for possible advances in earthquake prediction.

Research relying on various data sources such as GPS, ionospheric data and outgoing longwave radiation has improved earthquake prediction models. Gitis et al. (2021) demonstrated the value of seismological data, while Zhai et al. (2020) investigated thermal anomalies using non-seismic time series data, highlighting the multidisciplinary nature of earthquake research.

The work of Hsu and Pratomo (2022) on predicting early ground accelerations using long short-term memory (LSTM) neural networks demonstrated the usefulness of models that capture order dependence in seismic waves. This is consistent with the approach of using machine learning techniques to calculate earthquake events over time.

References:

¹ Improving Los Angeles Earthquake Prediction Accuracy Using Machine Learning – Cemil Emre Yavas et al. – Nature Scientific Reports – October 18, 2024 – – OPEN ACCESS