detection, Risk assessment, Early warning systems, Prediction models, Data analysis.

Earthquakes are among the most devastating natural disasters, often resulting in widespread destruction and loss of life. Accurate earthquake prediction remains a key focus within seismology and disaster preparedness, as even marginal improvements in forecasting could significantly enhance community safety and disaster readiness. In recent years, advancements in Machine Learning (ML) have opened up new avenues for analyzing large datasets and identifying complex patterns within seismic data, suggesting promising applications for earthquake prediction. This study leverages ML techniques to predict earthquake magnitude and likelihood in a region with a high frequency of seismic activity [1], [2]. Using the earthquake dataset, earthquake events from 1966 to 2007, we apply four ML models linear regression, Support Vector Machine (SVM), Naive Bayes, and Random Forest to identify key patterns and trends in earthquake occurrences. By examining factors such as magnitude, location, depth, and seismic station data, we aim to assess each model’s predictive accuracy, ultimately identifying the most effective algorithm for this specific application [3]. Our research contributes to earthquake prediction by comparing model performance through Mean Squared Error (MSE) and R² scores. Additionally, we explore the potential for enhancing predictive accuracy by integrating real-time data into these models, which could further support early warning systems and preparedness measures [4]. The findings from this study underscore the importance of selecting suitable ML models for specific seismic characteristics, with implications for the ongoing development of reliable, data-driven earthquake prediction tools [3].

The evolution of earthquake prediction research began with foundational techniques focusing on seismic activity monitoring and geophysical measurements. Anderson et al. [4] emphasized microseismic patterns as early warning signs, while Smith et al. [5] advanced the field with geodetic measurements, including GPS-based ground deformation tracking near tectonic fault lines. Lee et al. [6] introduced the study of electromagnetic emissions and radon gas anomalies as precursors, though these indicators often lacked consistency.

A significant shift occurred in the 2000s with Gupta et al. [7] utilizing historical earthquake data to develop probabilistic models. Despite these advancements, challenges in precision and scalability persisted, prompting researchers to explore computational approaches. Zhou et al. [8] demonstrated the potential of supervised learning models, while Chen et al. [9] applied unsupervised learning for anomaly detection in seismic data. The introduction of deep learning by Zhang et al. [10] utilizing Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs), significantly improved the analysis of seismic waveforms and feature extraction.

Hybrid models gained traction, with Patel et al. [11] combining ML and geophysical models for robust predictions. Liu et al. [12] addressed data scarcity through adaptive augmentation, while Park et al. [13] tackled model interpretability using advanced visualization techniques.

In 2024, Wu et al. [8] optimized real-time data streaming for deep learning frameworks, Kim et al. [14] integrated attention mechanisms into FasterNet, enhancing minor anomaly detection, and Singh et al. [15] developed adaptive models for regional tectonic variability. Tan et al. [16] documented these advancements, addressing challenges like data quality, computational constraints, and model scalability. This progression highlights the transition from traditional methods to sophisticated, data-driven techniques, steadily improving accuracy, real-time applicability, and disaster mitigation.

This study introduces a comprehensive machine-learning framework to predict earthquake magnitudes by leveraging critical seismic parameters, including latitude, longitude, depth, and the number of stations reporting the event. Accurate magnitude prediction is essential for disaster management and preparedness, as it enables authorities to take timely measures to mitigate potential damage and enhance public safety.

The proposed framework comprises several stages: data preprocessing, feature engineering, model training, and evaluation. The process begins with the meticulous cleaning and preparation of raw earthquake datasets. This involves handling missing or inconsistent data, removing outliers, and normalizing the dataset to ensure compatibility with ML models. Feature engineering follows, where the most relevant seismic parameters are selected and transformed to enhance the predictive capability of the models. This stage may also include creating additional features derived from the primary parameters to capture complex relationships in the data. Various ML algorithms are explored within the framework to identify the most effective approach for magnitude prediction [18]. These include linear regression for understanding linear dependencies, SVM for handling non-linear patterns, naive Bayes for probabilistic modeling, and random forest regression for capturing intricate relationships in the data through ensemble learning. Each model undergoes rigorous training using historical earthquake data, ensuring that the models can generalize well to unseen scenarios. The evaluation stage is a critical component of the framework, as it assesses the performance and reliability of the predictive models. A diverse set of metrics is employed to provide a holistic view of model effectiveness, including R-squared to measure the proportion of variance explained, MSE to quantify prediction errors, and accuracy to evaluate overall performance. Additionally, confusion matrices are used to analyze classification models where applicable, providing insights into specific areas for improvement. The framework's modular design ensures flexibility, allowing it to be adapted or extended as new data becomes available or as additional features are identified. By integrating robust preprocessing techniques, advanced feature engineering, and diverse ML approaches, this framework aims to deliver high accuracy in predicting earthquake magnitudes. Such advancements can significantly enhance disaster response strategies, resource allocation, and public safety planning.

The experiment is designed to evaluate the performance of several machine-learning models for predicting earthquake magnitudes. The dataset contains multiple features such as latitude, longitude, depth, number of stations, and the earthquake magnitude. The models are trained using a training set and evaluated using a test set to ensure unbiased performance.

Data preprocessing: The raw data is cleaned to remove missing or irrelevant entries. The Date and Time columns are merged into a single datetime index, and irrelevant columns are dropped.

Model selection: The following models are employed in the experiment:

I. Linear regression: A simple approach to model the relationship between the features and the magnitude. II. SVM is used to capture non-linear relationships and evaluate performance.

III. Naive bayes: Applied to predict earthquake categories based on magnitude ranges.

IV. Random forest regressor: A robust ensemble method to model the relationship between features and the magnitude.

V. Evaluation metrics: We use R-squared, MSE, and accuracy (for classification tasks) to assess the models' performance. The confusion matrix and classification report are also utilized for Naive Bayes to assess its performance in predicting categories.

VI. Environment: The experiment uses Python with libraries like Scikit-learn, Pandas, and Matplotlib to manipulate and visualize data.

Linear Regression: while linear regression is a fundamental approach, its performance in this context highlights the challenges of using a simple model for a complex problem like earthquake magnitude prediction. The relatively high MSE indicates that the model fails to provide highly accurate forecasts. However, it still offers a basic understanding of the relationship between the features and the target variable.

In this study, we applied ML models to predict earthquake magnitude using the earthquake dataset. Our findings reveal that the Random Forest model outperformed other methods, with the lowest MSE and highest R² value, underscoring its capability to capture intricate patterns in seismic data. This suggests ensemble models, particularly Random Forest, hold promise for earthquake magnitude prediction and Random Forest Classifiers for better risk assessment and prediction coordinates more prone to earthquakes. Future research might expand on this work by incorporating real-time seismic data and additional geological features. This could lead to more robust and responsive early warning systems, ultimately enhancing community preparedness and disaster mitigation efforts.

The data used in this study were sourced from the publicly accessible earthquake dataset. This dataset was utilized in its original form without any additional data generation. No new data were created during this study, and all relevant data can be accessed freely via the provided source.

The authors declare no conflict of interest. Funders played no role in the design of the study, data collection, analysis, or interpretation, nor in the writing of the manuscript or the decision to publish the results.

This section provides additional technical details about the experimental setup for evaluating the performance of machine learning models.

Data preprocessing: The dataset used in this study was cleaned to remove any missing or irrelevant entries. The following preprocessing steps were applied:

I. Handling of missing values through imputation techniques.

II. Convert the date and time fields into a single datetime index to facilitate time-based analysis.

III. Feature scaling for certain variables, such as depth and magnitude, to ensure uniformity in model inputs. Model hyperparameters: For each machine learning model used in the experiment, the following hyperparameters were tuned:

I. Linear regression: No specific tuning is required, as it is a simple linear model.

II. Svm: The kernel function was chosen as the Radial Basis Function (RBF), with hyperparameters C and γ adjusted using grid search.

III. Naive bayes: The Gaussian Naive Bayes model was employed, with no specific tuning.

IV. Random forest regressor: The number of trees was set to 100, and the maximum depth of the trees was varied between 5 and 15.