Introduction
From ancient times, following natural disasters, mankind has sought solutions to prevent or control these disasters. Earthquakes are one of these natural disasters that have always caused financial and human losses. The time, place, and magnitude of the earthquakes are the three main parameters of the earthquake assessment. To control and minimize its adverse effects, a good estimate of their values ​​should be available (Arjamand et al., 2015). Earthquake is a natural hazard with sudden shaking of the ground, which is of particular importance for human societies in terms of magnitude and area of destruction, unpredictability, and occurrence in a very short time (Gholami and Shokuhi Bidandi, 2023). For earthquake prediction, it is necessary to consider the time, place, magnitude, probability, and reason for its occurrence. The purpose of earthquake prediction is to help disaster management organizations prepare for earthquakes. When a strong earthquake is predicted, disaster management organizations should be alerted to take preventive measures. In disaster preparedness, decisions and activities focus on damage prevention. Various earthquake prediction techniques have been applied to mitigate damages (Bhatia et al., 2023). 
Shahrood, a district of Khalkhal county, Ardabil, Iran, is an area vulnerable to earthquakes due to the existence of geological formations with heterogeneous resistance, being located on an alluvial plain with less resistance than the thick bedrock, and being surrounded by numerous faults. Also, due to being located in the folded zone of Western Alborz and Azerbaijan, different faults in the region have been formed, the most important of which are Talash fault, Neor fault, and Shahooud Kalur fault. This research aims to predict the magnitude of possible earthquakes in the Shahrood district of Khalkhal County using artificial neural networks.

Methods
This research used various data and information, including the geological map of Rezvanshahr, Khalkhal, and Masuleh counties with a scale of 1:100,000, from which the map of faults and geological formations was extracted. ALOS-PALSAR digital elevation model with a spatial resolution of 12.5 m and Sentinel-2 satellite images were used to identify geological features. For finding different characteristics of earthquakes, the 30-year seismic data from the Institute of Geophysics of the University of Tehran and the global seismic data from the US Geological Survey were used, which included date of occurrence, magnitude, focal depth, and geographic coordinates. The selection of a 30-year period was due to the existence of complete data. For the period of more than 30 years, the data were mostly incomplete. All the data used for the preparation in the first stage were entered into the ArcGIS software and preliminary processing was done on all neural layers. After the preprocessing, the modeling is done using SPSS Modeler software.

Results 
For modeling the input data for the network, the data was split into two sets; 70% for training and 30% for testing. Five criteria were determined for predicting earthquake magnitude: Earthquake focal depth, magnitude of previous earthquakes, distance of the faults from the occurrence points, location of previous earthquakes, and fault length. After designing the model and configuring the network, the data was processed, and a multilayer perceptron (MLP) artificial neural network was created with 5 input neurons, 3 hidden neurons, and 1 output neuron. This model achieved an accuracy of 98.2%, indicating its high precision in processing data for predicting earthquake magnitude. The results indicated that, in the MLP model, the highest effect (0.83) was related to the magnitude of previous earthquakes, followed by the distance of the faults (0.58) and the focal depth of earthquakes (0.42). The lowest effects were related to the location of previous earthquakes (0.25) and fault length (0.18). The MLP model was finally found to have high validity and reliability in both training and testing phases.

Conclusion
The prediction results showed that earthquakes with a magnitude of 1-3 on the Richter scale are very likely to occur and the possibility of their occurrence on faults, especially the Kalur fault, is very high (70%). Although earthquakes with this magnitude have low risk, their occurrence in shallow depths and close to residential areas can cause mortality and financial loss. Also, earthquakes with a magnitude of 4-6 on the Richter scale have a moderate probability of occurrence (26%). These earthquakes can cause a lot of damage in the region. Moreover, the likelihood of earthquakes with a magnitude of 7-10 on the Richter scale in the area is 4%, indicating a low probability. In future studies, it is recommended to use machine learning models to better and more accurately predict earthquakes in the studied area.

Ethical Considerations

Compliance with ethical guidelines
All ethical principles were observed in this study.

Funding
This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

Authors' contributions
The authors contributed equally to preparing this paper.

Conflicts of interest
The authors declared no conflict of interest.

Acknowledgments
The authors would like to thank the National Center for Seismology of Iran and the Institute of Geophysics of the University of Tehran, for providing important data and information to carry out this research.




References