Scalable earthquake magnitude prediction using spatio-temporal data and model versioning

Abstract

Earthquake magnitude prediction is critical for natural calamity prevention and mitigation, significantly reducing casualties and economic losses through timely warnings. This study introduces a novel approach by using spatio-temporal data from seismic records obtained from the Indian government seismology department and weather data sourced via VisualCrossing to predict earthquake magnitudes. By integrating environmental and seismic variables, the study explores their interrelationships to enhance predictive capabilities. The proposed framework incorporates a machine learning operations (MLOps)-driven pipeline using MLflow for automated data ingestion, preprocessing, model versioning, tracking, and deployment. This novel integration ensures adaptability to evolving datasets and facilitates dynamic model selection for optimal performance. Multiple machine learning algorithms, including Gradient Boosting, Light Gradient Boosting Machine (LightGBM), XGBoost, and Random Forest, are evaluated on dataset sizes of 20%, 35%, 65%, and 100%, with performance metrics such as Mean Absolute Error, Mean Squared Error, Root Mean Squared Error, and R². The results reveal that Gradient Boosting performs optimally on smaller datasets, while LightGBM demonstrates superior accuracy with larger datasets, showcasing the pipeline’s flexibility and scalability. This research presents a scalable, robust, and resilient solution for earthquake magnitude prediction by combining diverse data sources with a dynamic and operational MLOps framework. The outcomes illustrate the potential of integrating advanced machine learning techniques with lifecycle management practices to enhance prediction accuracy and applicability in real-world seismic scenarios.

Introduction

The prediction and mitigation of natural calamities, such as earthquakes, present some of the most critical challenges in modern science and engineering. Earthquakes are complex natural phenomena that occur as a result of sudden energy release in the Earth’s crust¹, leading to ground shaking and potentially catastrophic impacts on human life and infrastructure². The monumental 2001 Bhuj earthquake, a devastating event in Gujarat, India, starkly demonstrated the severe impacts of natural calamities on human populations and economic stability. Occurring on January 26th, the earthquake resulted in over 13,800 fatalities, inflicted injuries on more than 167,000 people, and led to extensive destruction of homes and infrastructure. The earthquake caused significant economic losses, with an estimated financial impact of approximately $5 billion USD. This tragedy highlighted the urgent need for advanced earthquake prediction methods to mitigate future losses and protect vulnerable communities. Accurate prediction of earthquake events, including their magnitude and occurrence, remains elusive due to the non-linear and chaotic nature of the underlying processes. While significant progress has been made in understanding seismic activity, the development of reliable prediction models continues to be a topic of intense research³. Machine learning (ML) offers a promising approach to enhance earthquake prediction capabilities by exploiting patterns and relationships within large-scale datasets that traditional physical models may not fully capture.

The expansion of data-driven techniques to natural disaster prediction has been propelled by the recent availability of large amounts of environmental and geospatial (location-based) data. The adoption of machine learning has enabled modeling complex spatio-temporal interactions, allowing the extraction of minor signals that would be missed by conventional methods⁴. Incorporating data-driven models that can analyze various factors (atmospheric conditions, geological signatures, historical seismic records) into more complex modeling allows for greater accuracy of the likelihood of future earthquake magnitude concerning one or many covariates at any given time. Despite the prior benefits, large challenges remain around data quality and availability, generalizability of models, and recalibration to new data⁵.

One of the primary challenges in earthquake prediction lies in the inherent variability of seismic events⁶ across different geographical regions. Geological characteristics vary significantly from one area to another⁷, making it difficult to develop a universal model that performs well globally. Furthermore, the lack of sufficient high-quality labeled data poses a limitation for training sophisticated machine learning models⁸. These limitations necessitate innovative approaches to handle data scarcity and model adaptation. The adoption of machine learning operations (MLOps)⁹ practices in the workflow can address some of these challenges by facilitating the continuous integration, deployment, and monitoring of models¹⁰. MLOps introduces a systematic approach to managing the machine learning lifecycle¹¹, including automated data processing, model retraining, versioning, and reproducibility, thereby enhancing the robustness of predictive systems^10,12.

The objective of this study is to overcome the difficulties mentioned above by creating a predictive model for earthquake magnitude based on machine learning techniques, combining seismic and environmental data. Seismological data are downloaded from the seismology department of the Government of India (seismo.gov.in), which gives information about the epicenter of an earthquake, its latitude/longitude, depth, as well as magnitude. On top of that meteorological information from visualcrossing.com for weather: temperature, humidity, and precipitation; wind speed (average), atmospheric pressure, and solar radiation, When the model is trained based on these datasets, many scientific questions can be answered with a rich dataset from diverse sources, all of which use geological and environmental factors for identification that may influence seismic risk.

We developed a comprehensive machine learning workflow, incorporating MLOps principles and using MLflow as the core framework for managing the experiment lifecycle. The integration of MLflow facilitates the tracking of experiments, logging of model parameters and performance metrics, and deployment of models into production environments. The proposed framework ensures that the models remain up-to-date with new data and allows for continuous improvement over time¹³. Several state-of-the-art machine learning algorithms, including LightGBM¹⁴, XGBoost¹⁵, Gradient Boosting¹⁶, Random Forest¹⁷, and Linear Regression¹⁸, were evaluated to identify the most effective predictive model for earthquake magnitude. These models were trained on different subsets¹⁹ of the data (20%, 35%, 65%, and 100%) to assess their robustness and ability to generalize when data availability is limited.

For evaluation, we also use different types of parameters, like MAE (mean absolute error), MSE (mean squared error), MAE (mean absolute error), RMSE (root mean square error), and R². We chose this mixture of metrics to make sure we could properly test their specific kinds well: error dispersion, and explained variance on prediction (predictive power), as they tackle various parts from the puzzle²⁰. This multifaceted evaluation helps provide a more comprehensive understanding of how models will perform with different sizes, as well as determine if the data is diverse enough to adapt for generalization. The primary contributions of this research are as follows:

• 1. Novel MLOps-driven framework to build an end-to-end MLOps pipeline for earth-quake prediction with automated ingestion, data preprocessing, model training versioning, and deployment using MLflow. The pipeline handles the nonstationary nature of seismic activity, allowing model retraining as new data arrives in practice.

• 2. An innovative and comprehensive comparative analysis of different machine learning algorithms (LightGBM, XGBoost, Gradient Boosting, Random Forest and Linear Regression) against multiple dataset sizes (20%, 35%, 65% and 100%). Evaluation of performance measures: Mean absolute error (MAE), Mean squared error (MSE), Root mean squared Error (RMSE) and R². Our analysis shows that Gradient Boosting yields consistently good performance for the smallest data sizes, XGBoost has robustness to growing data size in the middle range, and that Light-GBM is able to fit well with datasets of full size. This adaptive model selection, based on data size and quality, sets a new benchmark for earthquake prediction systems.

• 3. A detailed analysis of specific weather patterns associated with elevations in earthquake hazard through an in-depth examination of settings using these same seismo-meteorological mechanisms Additionally, their work may aid in the formulation of new predictive models by showing how atmospheric phenomena can be precursors to seismic events.

• 4. A VERY real-world demonstration of sensitivity in availability and robustness of modeling performances. This work offers an out-of-the-box utility of accessing scalable machine learning techniques in the field of natural calamity prediction, especially when there is no data available to fit traditional models such as Linear Regression to infuse non-linearity.

• 5. Development of a new dynamic model selection framework within the MLOps pipeline, allowing to efficiently detect the most suitable algorithm and deploy it on demand based on the data properties on the fly. This ability is a major development for optimal performance in different seismic prediction situations.

The remainder of this paper is structured as follows: Section “Related work” reviews related work on machine learning applications in earthquake prediction and the use of MLOps frameworks in scientific research. Section “Data summary” details the data summary. Section “Methodology” details the methodology, including data preprocessing, model training, and evaluation procedures. Section “Result and discussion” presents the results and discussion, highlighting key findings and their implications for earthquake prediction. Finally, Section “Conclusion” concludes the paper and suggests directions for future research.

Related work

Prediction of natural calamities, e.g. earthquakes have been the focus area, and a lot of research is being done to improve the accuracy using spatiotemporal data and machine learning techniques. More recent efforts have incorporated environmental considerations in combination with seismic data to gain a more accurate accounting of the vast number of processes that contribute towards triggering any given occurrence. Moreover, growing acceptance of MLOps practices for predictive model lifecycle management guarantees the robustness of these models in real-world scenarios.

This study²¹ delves into the implementation of serverless computing within machine learning systems, focusing primarily on the aspect of covariate drift detection. Their research demonstrates how serverless architectures can not only facilitate but also enhance the scalability and cost-efficiency of monitoring machine learning models in production environments. The approach promises significant advancements in handling large-scale data without the need for extensive infrastructure management. However, the study is predominantly theoretical and primarily validated in simulated environments, which brings into question its effectiveness in real-world applications. This gap suggests a need for additional empirical research to ensure that the proposed serverless computing framework can operate effectively under diverse and dynamic real-world conditions, thereby proving its practical viability across various industries. This study²² look at the forward-thinking use of Geographic Information Systems (GIS), along with machine learning algorithms, like Decision Trees and Random Forests, can improve the evaluation of geological disaster susceptibility. The objective of this synergy is to make improvements in the predictability and spatial analysis for effective disaster response through counter-disaster management. The study results can be useful in mapping disaster-prone areas using GIS technology for improved pre-paredness and resource allocation, especially during geological events. Nevertheless, extensive and precise geographical data become the main constraints of this model. However, in areas where data is sparse or unreliable, the model performs poorly and is not very useful. This limitation is the most critical, especially relevant for less developed areas where assessments of geological disaster susceptibility are perhaps urgently in demand. This gap can be closed by either improving the methodologies for data collection or building models that provide robust outputs from low and noisy inputs. In response to the aforementioned challenges,²³ calls for MLOps in continuous development pipelines for artificial intelligence. This is a more efficient way of managing the AI model’s development, deployment, and maintenance lifecycle. The study further signals a need for centralized tools and agreement on best practices customized for AI development that are transferrable across disparate technologies. Absence of standards and dedicated tools for AI lifecycle management indicates a huge void in the domain that adversely impacts the smooth adoption of MLOps practices across organizations. The other huge gap is that we need generic standards and tools that are designed based on the specific nature of AI operations to be more widely applicable, which will help in faster adoption of MLOps practices.

This study²⁴ describe essential features for an MLOps platform for industrial metal AM use. The development platform aims to simplify the management of machine learning models to make manufacturing processes more efficient. This type of platform enables better scalability of ML applications, as well as the capacity to keep adjusting models as new data becomes available. That said, integrating such complex platforms with aging manufacturing systems not originally designed to support advanced ML functionalities is one of the most significant challenges. The existence of legacy systems in many manufacturing facilities that may necessitate costly upgrades/replacements to facility and service-based systems is a financial/logistical barrier that could hinder the implementation of novel ML technologies based on traditional capabilities in almost all manufacturing environments.

This study²⁵ applies stacking ensemble learning methods by combining different machine-learning techniques to improve the accuracy prediction of power outage duration during typhoon events. The ensemble model, which is a composite of multiple algorithms, can predict with great accuracy. While the model is effective, its complexity and high computational requirement might limit where it can be deployed especially in areas with limited access to advanced technological infrastructure. This restriction is significant since the capacity to forecast power outages reliably is in high demand, especially for less developed areas more vulnerable to extreme weather threats.

This study²⁶ apply machine learning operations (MLOps) for automating and optimizing freight rail operations, with emphasis on real-time delay predictions. MLOps has a huge impact on operational efficiency and is quite representative of ML transforming traditional industries. But the paper also points out that these predictions have not yet been thoroughly validated against performance in different geographic or infrastructural settings. This is a critical research gap, as we need to know how the parameter performs under various operational conditions through proper testing, without which deployment may not be so reliable and adaptable.

This study²⁷ proposed an extremely brief urban flood forecasting model based on K-means clustering and a Bayesian optimized deep learning approach, focusing on the special environmental conditions of coastal cities and so forth. Such predictions are central to any strategy for flood management and response, thus benefiting from the precision and rapidity of this approach. This reliance on precise environmental data for a high-performing model is a critical weakness since in many locations environmental data may not exist or may be of poor quality. However, this reliance on quality data limits the domain of application for the model to areas that may be exactly where better models are sought, signifying a demand for models that generalize with suboptimal input–output combinations.

This study²⁸ qualitatively investigates how manufacturing domains can influence the operationalization of machine learning models and provide guidelines for integrating MLOps into such environments so that ML algorithms can be continuously delivered and managed in manufacturing systems. The potential of MLOps to revolutionize manufacturing operations by facilitating a more agile and incremental application of machine learning is highlighted in their research. On the other hand, they also emphasize a major research gap: the lack of empirical evidence to support the theoretical benefits of MLOps in manufacturing. This gap suggests that more practical, industry-specific research is needed to validate the proposed benefits and to adapt MLOps practices to meet the unique challenges of the manufacturing sector.

Several recent studies have demonstrated the applicability and effectiveness of the DAS magnitude scale (Mwg) in seismic research and hazard assessments across various tectonic settings. In their foundational work, Das and Meneses proposed scaling relations for energy magnitudes, establishing Mwg as a robust alternative for quantifying seismic energy release²⁹. This approach was extended by Das et al. in the development of a unified earthquake catalog for Northeast India, where the use of Mwg enabled enhanced consistency and reliability of magnitude estimates across the region³⁰. In a similar vein, Boudebouda et al. applied a refined version of the Mwg scale based on robust regression techniques to create a comprehensive seismic catalog for Northern Algeria³¹. Most recently, Das et al. integrated Mwg into a probabilistic seismic hazard evaluation for northern Chile, combining it with GNSS strain data to improve the precision of hazard modeling³². These studies collectively underscore the versatility and growing adoption of the Mwg scale in modern seismological analyses.

Despite significant advancements in earthquake prediction using spatio-temporal data, existing studies often face limitations in data quality, integration of diverse environmental factors, and practical deployment challenges. These gaps highlight the need for further research to develop more robust, scalable, and adaptable predictive models. Our research aims to address these issues by enhancing model accuracy and operational efficiency in real-world natural calamity forecasting scenarios.

Data summary

In this study on earthquake prediction using spatio-temporal factors, two main datasets were employed: earthquake records obtained from the Indian government seismology department (seismo.gov.in) and weather data sourced from VisualCrossing.com. These datasets form the basis for investigating the relationships between seismic activity and environmental conditions over time.

Earthquake data

The earthquake dataset, obtained from the official Indian government seismology site (seismo.gov.in), provides core details of seismic events across different geographic locations in India. It includes key features such as latitude (Lat) of the earthquake’s epicenter, longitude (Long) of the earthquake’s epicenter, depth of the earthquake in kilometers which reflects its distance beneath the Earth’s surface, the Richter scale magnitude representing the energy released by the earthquake, and the descriptive location of the earthquake’s occurrence.

This dataset captures the spatial attributes (latitude, longitude, and depth) and the earthquake’s intensity (magnitude), essential for understanding the geographic distribution and severity of seismic activity.

Weather data

To complement the earthquake dataset, a weather dataset was downloaded from VisualCrossing.com. This dataset provides historical and environmental conditions for corresponding latitude and longitude values in the earthquake dataset. Key features include various data categories:

Temporal Data: Includes datetime for the specific date and time of recorded weather data, along with Sun Cycle Data such as sunrise and sunset timings.

Temperature Data: Encompasses tempmax, tempmin, and temp which record the maximum, minimum, and average temperatures (in degrees Celsius). It also includes feelslikemax, feelslikemin, and feelslike for the maximum, minimum, and average apparent temperatures, and dew for the dew point temperature, indicative of humidity levels.

Precipitation and Snow Data: Contains precip for the total precipitation amount, precipprob and precipcover for the probability and cover of precipitation, preciptype for types of precipitation like rain or snow, and snow and snowdepth for the amount of snowfall and depth of snow on the ground.

Wind and Atmospheric Data: Includes windgust and windspeed for gust and sustained wind speeds, winddir for wind direction in degrees, and sealevelpressure for atmospheric pressure at sea level, relevant for weather system analysis.

Visibility and Cloud Data: Covers visibility for distance visibility, crucial for assessing environmental clarity, and cloudcover for the percentage of sky obscured by clouds.

Solar and UV Data: Provides solarradiation and solarenergy for solar radiation and solar energy levels, and uvindex for the ultraviolet index, relevant for understanding solar exposure levels.

Severe Weather Indicators: Includes severerisk for potential risk level for severe weather conditions.

This dataset enriches the earthquake data with diverse environmental factors, enabling analysis of how weather and atmospheric variables interact with seismic activities. The temporal alignment and extensive coverage of both datasets allow for the exploration of potential correlations between natural calamities and environmental conditions, providing a foundation for predictive modeling in this domain.

Methodology

The methodology for this research is structured into several key steps: Magnitude Scale Transformation, Seismic Data Analysis, data preprocessing, machine learning model training, model versioning and tracking using MLOps (implemented with MLflow), and continuous model evaluation with new data integration.

Magnitude scale transformation

To ensure methodological rigor and accurate representation of earthquake magnitudes, we standardized all input magnitude values using a multi-step transformation pipeline aligning with recent scientific advancements. The initial dataset reported earthquake magnitudes in the Local Magnitude (ML) scale. However, ML is known to saturate for larger events and can introduce scale inconsistencies. To address this, the magnitude values were transformed to the Das magnitude scale ($M_{wg}$) using a systematic conversion process derived from recent seismological studies^33,34,35.

By applying this conversion, we initially standardized all earthquake magnitudes to the $M_{w}$ scale, as shown in Eq. (1), to address the saturation limitations of ML for large-magnitude events. However, recent investigations have highlighted that the $M_{w}$ scale, despite its improvements over ML, introduces significant biases and lacks reliability for events below magnitude 7.5^33,34. These shortcomings are attributed to the use of surface wave amplitudes in $M_{w}$ calculations, which may not accurately capture the earthquake source energy, especially in varied tectonic settings.

To enhance methodological accuracy and ensure a more globally consistent magnitude representation, we adopted the Das Magnitude Scale ($M_{wg}$), which has been developed and validated using over 25,000 global seismic events³³.

First, ML values were converted to the Moment Magnitude ($M_{w}$) scale using the empirical relation:

$$M_{w} = 0.67 \times ML + 0.33$$

(1)

The Seismic Moment Magnitude ($M_{0}$) was computed from the converted $M_{w}$ values:

$$\log M_{0} = \left( {M_{w} + 10.7} \right) \times \frac{3}{2}$$

(2)

Here, M₀ is expressed in dyn.cm, and this formulation provides a physically consistent estimate of energy release during seismic events.

Finally, the Das magnitude scale ($M_{wg}$) was computed:

$$M_{wg} = \frac{{\log M_{0} }}{1.36} - 12.68$$

(3)

While the Moment Magnitude ($M_{w}$) scale is widely adopted due to its ability to overcome ML saturation, it has critical limitations, particularly for earthquakes below $M_{w}$ 7.5. As shown in recent studies^33,34, $M_{w}$ tends to underestimate energy release in small-to-moderate events due to its reliance on surface wave amplitudes and regional calibrations. These shortcomings compromise its global applicability and accuracy in machine learning tasks requiring consistent magnitude representation.

To address this, we adopted the Das Magnitude Scale ($M_{wg}$), originally proposed in³⁵. $M_{wg}$ is derived from global seismic data and calibrated using body wave and moment relationships, providing superior energy correlation and reliability across all depths and magnitude ranges. Its use ensures that our model training reflects physically consistent and globally validated seismic characteristics.

Seismic data analysis

In the exploratory analysis of seismic data, various graphical representations were used to gain insights into the distribution, correlation, and temporal patterns of earthquake occurrences. The 3D spatial distribution plot in Fig. 1 illustrates the distribution of earthquake magnitudes across different geographic locations, with longitude, latitude, and magnitude plotted on the x, y, and z-axes, respectively. This plot reveals spatial clustering of seismic events, indicating regions with higher frequencies of seismic activity and varying magnitudes.

Fig. 1

3D spatial distribution of earthquake magnitudes. The figure was generated using Python with Matplotlib (v3.9.2)³⁶ and Seaborn (v0.13.2)³⁷.

The feature correlation heatmap in Fig. 2 provides a matrix of correlation coefficients among key seismic variables, including latitude, longitude, depth, and magnitude. This visualization highlights the relationships between these variables, showing a moderate correlation between depth and magnitude. Such correlations offer insight into potential predictive relationships and guide the selection of features for modeling. A histogram of depth shown in Fig. 3 depicts the frequency distribution of earth-quake depths. This plot shows that most earthquakes occur at shallow depths, with a steep drop-off as depth increases. This pattern suggests that shallow earthquakes are more frequent, which may influence their impact and relevance in predictive modeling. The box plot of magnitude by depth categories in Fig. 4 segments earthquake magnitudes based on defined depth ranges. This plot provides a comparison of magnitude distributions within each depth category, showing that deeper earthquakes generally have more consistent magnitudes, while shallower earthquakes exhibit a wider range of magnitudes, including more intense events.

Fig. 2

Feature correlation Heatmap (Seismic Data).

Fig. 3

Histogram of depth.

Fig. 4

Magnitude distribution by depth categories.

The pair plot in Fig. 5 is a multi-dimensional scatter plot that visualizes the pairwise relationships between latitude, longitude, depth, and magnitude. It shows the distribution of each variable along the diagonal and the interactions between variable pairs in the off-diagonal plots. This plot enables a quick visual examination of how features interact and distribute in relation to each other, identifying any visible trends or clustering in the data.

Fig. 5

Pair plot of seismic features. The figure was generated using Python with Matplotlib (v3.9.2)³⁶ and Seaborn (v0.13.2)³⁷.

The scatter plot of magnitude versus depth in Fig. 6 directly visualizes the relationship between earthquake depth and magnitude. It reveals that while there is a broad range of magnitudes across different depths, higher magnitudes tend to occur at shallower depths, though extreme depths occasionally show higher magnitudes.

Fig. 6

Scatter plot of magnitude versus depth.

This relationship is essential in understanding how depth may influence the severity of seismic events.

Finally, the yearly trend in earthquake occurrences shown in Fig. 7 provides a line plot showing the number of earthquakes recorded each year. This temporal analysis reveals fluctuations in seismic activity over time, which may reflect both natural seismic cycles and variations in detection capabilities or reporting. This trend analysis helps to identify any potential increase or decrease in seismic activity over recent years, informing the relevance and timeliness of predictive efforts.

Fig. 7

Yearly trend in earthquake occurrences.

These graphical analyses form the foundation for understanding the distributional characteristics of the seismic data and the relationships among variables, which are essential for building effective predictive models in this research.

Data preprocessing

Effective data preprocessing is crucial for improving model performance and reliability. Our preprocessing pipeline addresses missing values, outliers, noise reduction, and feature engineering to enhance data quality. Missing values in key attributes such as temperature and visibility were imputed using median and mode strategies to maintain dataset integrity. Outliers, particularly in temperature and precipitation, were identified using the interquartile range (IQR) method and removed to prevent skewed predictions.

Noisy data can obscure significant patterns essential for accurate predictions. To mitigate this, advanced filtering methods and noise-reduction algorithms were applied during preprocessing. Specifically, Empirical Mode Decomposition (EMD)³⁸ was utilized to decompose complex signals into intrinsic mode functions (IMFs), effectively separating noise from valuable seismic information. This process enhances the signal-to-noise ratio, allowing models to focus on relevant seismic patterns and improving the reliability of magnitude estimations.

To enhance the dataset’s predictive power, we performed feature engineering, extracting temporal attributes (month, day, year) and applying a logarithmic transformation to skewed variables like precipitation. Standardization was applied to numerical features such as temperature, wind speed, and humidity to ensure consistency across different scales. Additionally, Principal Component Analysis (PCA) was used to reduce dimensionality while preserving 95% of the data’s variance. To balance the dataset and improve generalization, Synthetic Minority Over-Sampling (SMOTE)^39,40 was applied to handle class imbalances in earthquake magnitudes.

These preprocessing steps ensure that the input data is clean, structured, and optimized for training machine learning models, leading to improved prediction accuracy and robustness.

Machine learning models

Various advanced machine learning algorithms are chosen for model training, such as LightGBM, XGBoost, Gradient Boosting, Random Forest and Linear Regression. These models can handle non-linear relationships, which are important due to the complex nature of natural calamity prediction. Each model’s performance was optimized by carefully tuning key hyper-parameters. For Linear Regression, we implemented it without regularization, focusing on its simplicity and interpretability. In the Random Forest model, we set the number of trees (n estimators) to 100, balancing computational efficiency and performance, and left the maximum depth (max depth) unrestricted to capture complex interactions. The Gradient Boosting model was configured with 100 boosting stages (n estimators), a learning rate (learning rate) of 0.1 to control each tree’s contribution, and a subsample (subsample) of 1.0, utilizing all training data per iteration. For XGBoost, we set 100 boosting rounds (n estimators), a learning rate (learning rate) of 0.1, and a maximum tree depth (max depth) of 6 to balance complexity and overfitting. The LightGBM model was configured with 100 boosting iterations (n estimators), a learning rate (learning rate) of 0.1, and a maximum depth (max depth) of 10 to prevent overfitting while capturing essential patterns. This meticulous hyper-parameter tuning was essential in developing a robust and accurate earthquake magnitude prediction framework.

The training process was conducted in four stages, progressively increasing the dataset size to evaluate model performance under varying data availability scenarios. The first run used 20% of the dataset (split into 70% training and 30% testing), simulating an early-stage learning scenario with limited historical data. The second iteration expanded the dataset to 35%, followed by 65% in the third iteration, representing moderate data accumulation over time. Finally, the fourth iteration utilized the full dataset (100%), ensuring models were trained on all available seismic and environmental records. This progressive strategy reflects real-world earthquake prediction conditions, where newer data continuously becomes available, allowing models to adapt dynamically. By analyzing performance across these dataset splits, we assess how models generalize under different temporal constraints, ensuring their robustness and scalability in earthquake prediction.

The performance of the models was evaluated using the following metrics:

• a. Mean Absolute Error (MAE):

  $${\text{MAE}} = \frac{1}{n}\mathop \sum \limits_{i = 1}^{n} \left| {{\text{y}}_{\mathrm{i}} - \hat{{\text{y}}}_{\mathrm{i}} } \right|$$

  (4)

  where ${\text{y}}_{{\text{i}}}$ is the true value and $\hat{{\text{y}}}_{\mathrm{i}}$ is the predicted value for the i-th instance, and n is the number of data points.

• b. Mean Squared Error (MSE):

  $${\text{MSE}} = \frac{1}{n}\mathop \sum \limits_{i = 1}^{n} \left( {{\text{y}}_{\mathrm{i}} - \hat{{\text{y}}}_{\mathrm{i}} } \right)^{2}$$

  (5)

  MSE calculates the average of the squared differences between predicted and actual values, penalizing larger errors more significantly than MAE.

• c. Root Mean Squared Error (RMSE):

  $${\text{RMSE}} = \sqrt {\left( {\frac{1}{n}\mathop \sum \limits_{i = 1}^{n} \left( {{\text{y}}_{\mathrm{i}} - \hat{{\text{y}}}_{\mathrm{i}} } \right)^{2} } \right)}$$

  (6)

  RMSE gives greater weight to larger errors compared to MAE.

• d. R-squared (R²):

  $$R^{2} = 1 - \frac{{\mathop \sum \nolimits_{i = 1}^{n} \left( {{\text{y}}_{\mathrm{i}} - \hat{{\text{y}}}_{\mathrm{i}} } \right)^{2} }}{{\mathop \sum \nolimits_{i = 1}^{n} \left( {{\text{y}}_{\mathrm{i}} - \overline{{\text{y}}}} \right)^{2} }}$$

  (7)

  where $\overline{\text{y}}$ is the mean of the true values. This metric explains the proportion of the variance in the dependent variable that is predictable from the independent variables. Together, these metrics provide a comprehensive assessment of model performance, and capture error magnitudes.

MLOps framework

To ensure seamless model deployment and management, we integrated machine learning operations (MLOps) using MLflow. MLOps automates key tasks such as experiment tracking, model versioning, and continuous deployment, making the earthquake prediction pipeline scalable and adaptable to new data.

The MLOps workflow follows a structured process: data ingestion, preprocessing, model training, evaluation, and deployment. MLflow Tracking logs model parameters, performance metrics, and dataset versions, enabling transparent model comparisons. Model Versioning^41,42 ensures that the best-performing model is always deployed while maintaining past iterations for reproducibility. Automated retraining is triggered when new seismic data becomes available^43,44, allowing models to remain up-to-date.

This framework enhances scalability, reproducibility, and automation, ensuring that earthquake prediction models remain robust and reliable over time. Future work will focus on optimizing real-time data streaming and computational efficiency.

MLflow: a tool for managing the ML lifecycle

MLflow is an open-source platform for managing the machine learning lifecycle⁴⁵, including experimentation, reproducibility, and deployment⁴⁶. With a set of pieces to record experiments, reproduce experiments, and use them in deployment, it supports important functionalities: tracking, packaging projects, deploying models, and versioning^47,48.

MLflow Tracking allows data scientists to log and then query experiment parameters, metrics, and results. We also used MLflow to monitor the performance of various natural calamity prediction models along iterations of spatio-temporal data. To find the best model performance, metrics like MAE, MSE, RMSE, and R² were logged. The logged tracking information can be represented as follows:

$${\text{Tracking Logs}} = \{ {\text{hyper-parameters, model parameters, performance metrics, artifacts}}\}$$

(8)

MLflow projects provide a standardized format for packaging code, ensuring experiment reproducibility. For this research, the entire workflow for natural calamity prediction was encapsulated into MLflow projects, enabling other researchers to rerun experiments with identical configurations and reproduce results reliably⁴⁹.

MLflow Models standardize the model packaging process, facilitating easy deployment across various environments, including REST APIs and cloud platforms. By packaging models in the MLflow Models format, we ensured seamless deployment and monitoring of our calamity prediction models in production environments⁵⁰.

MLflow Registry is important to manage the lifecycle of machine learning models. With it, teams can register models, deploy them into production, and manage their revision history. The MLflow Registry was used in this research work to store the versions of models and allow deployment of only those with the best performance for real-time prediction during calamities⁵¹.

MLflow was useful in tracking model performance and automating the deployment pipeline for our natural calamity prediction model. Using MLflow, we could still log all the iterations of our model having incrementally increased sizes of our training set (20%, 35%, 65%, and finally 100% of our data). This allowed us to track the performance from each version and find out which model performs better so it can be deployed in real-time.

For example, when new data arrived, MLflow enabled us to automatically retrain and redeploy the models. This process of retraining was automated so that the model would always have updated weather and disaster data in near real-time to make predictions effectively.

Advancements in MLOps pipeline

Our framework enhances MLflow’s capabilities by introducing dynamic model selection, automated hyper-parameter optimization, advanced data handling, and scalable deployment. It evaluates multiple machine learning algorithms like Linear Regression, Random Forest, Gradient Boosting, XGBoost, and LightGBM to adaptively select the most suitable model based on real-time data characteristics and performance metrics. Automated hyper-parameter tuning employs grid search and cross-validation to identify configurations that minimize prediction errors without manual intervention. The system ingests real-time data from seismic sensors and environmental monitoring, applying advanced feature engineering to capture critical spatio-temporal patterns and energy-depth interactions, ensuring models are trained on current data for improved responsiveness to seismic activity. Designed for scalability, the pipeline supports distributed computing and can be deployed across various platforms, including cloud services and on-premises infrastructure, accommodating diverse operational requirements and resources.

Algorithm for model versioning with MLflow

The following algorithm outlines the steps for managing the machine learning lifecycle using and MLflow.

Algorithm 1

Model Versioning with MLflow

Algorithm for prediction based on updated model

Once a model has been deployed, it is used to predict natural calamities. The following algorithm outlines the prediction process using the latest deployed model.

Algorithm 2

Calamity Prediction using Latest Model

Result and discussion

This section presents the results of evaluating five machine learning models LightGBM, XGBoost, Gradient Boosting, Random Forest, and Linear Regression on various subsets of the dataset (20%, 35%, 65%, and 100%) for natural calamity prediction. The evaluation focuses on key performance metrics such as mean absolute error (MAE), mean squared error (MSE), root mean squared error (RMSE), and R² shown in the Tables 1, 2, 3, and 4.

Table 1 Comparison of metrics of algorithms with 100% dataset used.

Table 2 Comparison of metrics of algorithms with 65% dataset used.

Table 3 Comparison of metrics of algorithms with 35% dataset used.

Table 4 Comparison of metrics of algorithms with 20% dataset used.

Performance degradation across dataset sizes

One of the main objectives of this research was to evaluate how each algorithm performed as the dataset size decreased. Tables 5 and 6 summarizes the error of each algorithm across different dataset sizes and quantifies the percentage degradation as the dataset size is reduced from 100 to 20%.

Table 5 Performance degradation across dataset sizes for RMSE.

Table 6 Performance degradation across dataset sizes for MSE.

Their results emphasize that Gradient Boosting is least impacted when it comes to the RMSE and MSE performance metrics, proving its great robustness to decreased dataset sizes. Thus it proves extremely useful in reducing errors and is less sensitive to the number of data points used to train, which is very effective for predictive tasks like the prediction of natural calamities. LightGBM, although it has a much worse MSE, is still more stable for the RMSE than any other model, including Random Forest, which shows huge degradation in both metrics. In contrast to the others, Linear Regression shows a specific trend of negative degradation in MSE and RMSE, slight improvements with less data; however, this could simply be due to its overall lower performance and simpler model dynamics that are not generally utilized for prediction tasks.

Mean absolute error (MAE) comparison across dataset sizes

Table 7 compares the mean absolute error (MAE) for each model across the four dataset sizes. The results reveal that both LightGBM and Gradient Boosting maintain low MAE values even when the dataset size is significantly reduced. Gradient Boosting and LightGBM demonstrate minimal changes in MAE, highlighting their superior ability to handle limited data without significant performance deterioration.

Table 7 Mean absolute error (MAE) comparison across dataset sizes.

In contrast, XGBoost and Random Forest experience moderate increases in MAE as the dataset size decreases, indicating that while they perform reasonably well, they are less reliable compared to Gradient Boosting and LightGBM. Linear Regression consistently exhibits the highest MAE across all dataset sizes, reaffirming its inadequacy for handling the complexity and variability inherent in natural calamity prediction tasks.

Overall, the models with the lowest MAE, such as LightGBM and Gradient Boosting, continue to demonstrate their effectiveness in managing the predictive complexity of the given task, even with reduced training data.

Root mean squared error (RMSE) comparison across dataset sizes

Table 8 compares the root mean squared error (RMSE) of each model across different dataset sizes, offering a clear view of the prediction errors’ spread. The analysis shows that Gradient Boosting and LightGBM maintain low RMSE values across all dataset sizes, making them highly suitable for natural calamity prediction tasks, where minimizing the variance in error is crucial.

Table 8 Root mean squared error (RMSE) comparison across dataset sizes.

On the other hand, XGBoost and Random Forest demonstrate moderate increases in RMSE as the dataset size decreases, but they still perform reasonably well, though worse than LightGBM and Gradient Boosting. Linear Regression, as expected, exhibits the highest RMSE values across all dataset sizes, further reinforcing its unsuitability for this particular task as it struggles to capture the complexities of the spatio-temporal features used in predicting natural calamities.

The models with consistently low RMSE, such as LightGBM and Gradient Boosting, further underscore their suitability for tasks that require both precision and robustness in predictions.

Figure 8 provides a grouped bar plot comparing the RMSE for each model across different dataset sizes. LightGBM and Gradient Boosting maintain the lowest RMSE across all dataset sizes, indicating their superior ability to minimize the spread of errors. This is crucial in applications like natural calamity prediction, where reducing prediction variance is critical for decision-making and planning. XGBoost and Random Forest show moderate increases in RMSE as dataset size decreases, but they remain competitive with only slightly higher RMSE values than LightGBM and Gradient Boosting. Linear Regression consistently exhibits the highest RMSE values, which worsen as the dataset size decreases, reinforcing its poor performance and lack of generalization capability for this problem. This further highlights the importance of choosing non-linear models in complex prediction tasks involving spatio-temporal data.

Fig. 8

RMSE comparison across dataset sizes: bar chart depicting the RMSE of the models across various dataset sizes, highlighting changes in error rates with dataset size.

Insights from MAE distribution across models

Figure 9 shows a box plot representing the distribution of mean absolute error (MAE) for each model across different dataset sizes. LightGBM and Gradient Boosting have the tightest MAE distribution, indicating that they are not only more accurate but also more consistent in minimizing absolute errors across varying dataset sizes. XGBoost and Random Forest show moderate variability in MAE but maintain lower error ranges compared to Linear Regression, which exhibits a wider MAE range. The outliers in Linear Regression highlight its instability, with a significant spread in MAE across datasets, reinforcing the notion that it is not an ideal candidate for this problem. The consistency observed in gradient-based models underlines their reliability in natural calamity prediction scenarios.

Fig. 9

Distribution of MAE across models: box plot showing the distribution of mean absolute error (MAE) across the models.

As it can be presented by the Figs. 10 and 11, although all algorithms did observe high error rates when only small datasets were available, Gradient Boosting and Light-GBM saw the least rapid increase in RMSE (and MSE) suggesting that they are probably more robust models under such data constraints.

Fig. 10

Line graphs illustrating the evolution of mean squared error (MSE) across varying dataset sizes for five predictive algorithms.

Fig. 11

Line graphs illustrating the evolution of root mean square error (RMSE) across varying dataset sizes for five predictive algorithms.

As shown in the graphs from Figs. 12 and 13, it can be verified that for all sizes of dataset Linear Regression yields higher RMSE and MSE when compared with other models such as LightGBM and Gradient Boosting which retain lower error levels even for smaller-sized datasets, suggesting high predictive accuracy power of these two models.

Fig. 12

Comparative bar graphs for MSE across different dataset sizes, segmented by algorithm.

Fig. 13

Comparative bar graphs for RMSE across different dataset sizes, segmented by algorithm.

As it can be seen by Fig. 14, Linear Regression has the lowest relative increase in errors, which is possibly due to the simpler representation of its model and thereby less sensitivity to reduced training data. But its absolute error is still quite high, even though the percentage increases are lower.

Fig. 14

Percentage degradation in RMSE and MSE as dataset sizes reduce from 100 to 20%.

Heatmaps in Figs. 15 and 16 provide an intuitive overview of the results across all algorithms (columns) and dataset sizes (rows), showing that we it always counted on the good performance of Gradient Boosting or LightGBM over any size, while high error rates for Linear Regression can be seen at reduced dataset sizes.

Fig. 15

Heatmaps of MSE values across various algorithms and dataset sizes.

Fig. 16

Heatmaps of RMSE values across various algorithms and dataset sizes.

Feature importance analysis

To understand the contribution of different features to earthquake magnitude prediction, we employ Permutation Importance, a model-agnostic technique that evaluates how the shuffling of a feature’s values affects the model’s performance. Figure 17 presents the ranked feature importance, showing that geospatial and seismic parameters play a crucial role in the prediction process. The results indicate that certain features have a higher predictive influence than others, providing insights into the factors most relevant for modeling earthquake magnitude.

Fig. 17

Feature importance plot using permutation importance. The permutation-based ranking showcases the significance of each feature in the earthquake magnitude prediction model.

Practical challenges in MLOps implementation

While our study demonstrates the feasibility of integrating MLOps into earthquake magnitude prediction, several practical challenges must be considered for real-world deployment. Real-time data streaming remains a critical hurdle, as ensuring seamless ingestion of seismic and environmental data with minimal latency requires robust data pipelines and scalable infrastructure. Computational costs pose another constraint, particularly when deploying machine learning models on resource-intensive cloud platforms or edge devices. Efficient model retraining strategies and optimized hyper-parameter tuning can help mitigate these expenses. Additionally, system reliability under operational conditions is essential, as model performance can degrade over time due to data drift or unexpected anomalies. Implementing automated monitoring, model retraining, and anomaly detection mechanisms within the MLOps pipeline can enhance system robustness. Future work should focus on optimizing these aspects to improve real-world applicability and scalability of earthquake prediction models.

In this paper, we have performed an analysis of different machine learning models (LightGBM, XGBoost, Gradient Boosting, Random Forest, and Linear Regression) over multiple dataset sizes (20%, 35%, 65%, and 100%) in predicting natural calamities using standard performance parameters such as MAE, MSE, (RMSE), (R²) Our visual and tabular analyses show that model robustness and adaptability are significantly different from one another, especially under limited data conditions.

The fact that performance degrades with fewer data is interesting, denoting the use of ensemble methods like LightGBM, XGBoost, or Gradient Boosting works to varying degrees, and they manage to withstand this degradation. Gradient Boosting works well for small datasets because it can work with sparse data, and XGBoost adapts nicely to medium-sized datasets. However, when compared with other models trained on the full dataset, LightGBM takes advantage of bigger volumes of data, proving to be both scalable and efficient. The results also highlight that due to the inherent variability in the real-world availability, data quality can fluctuate, implying that using the model that performed well in one case might not be true in the other case; therefore, we have shown how the selection of the best-performing model is important based on the available dataset size and so helps achieve realistic predictions.

Linear Regression consistently underperforms across all evaluated metrics, high-lighting its limitations in capturing the non-linear relationships inherent in spatio-temporal data. In contrast, LightGBM and Gradient Boosting maintain the lowest error rates (MAE, MSE, and RMSE) across most dataset sizes, reinforcing their robustness and suitability for complex predictive tasks such as natural calamity fore-casting. These findings provide valuable insights into the importance of dynamic model selection within an MLOps framework, enabling the deployment of the most effective models based on real-time data conditions. This approach not only enhances predictive accuracy but also ensures the scalability and adaptability of predictive systems in the field of natural calamity forecasting.

Conclusion

Natural calamity prediction is a broad and comprehensive research area, and the work reported here addresses earthquake magnitude prediction using spatio-temporal data. This study focuses on the importance of data availability with respect to model performance by assessing various machine learning models using LightGBM, XGBoost, Gradient boosting, Random Forest, and Linear regression with datasets of 20%, 35%, 65%, and 100%. Evaluate models based on metrics: Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and (R²) The results indicate that Gradient Boosting has a strong performance on smaller datasets, XGBoost achieves the best scores with medium-sized datasets, while LightGBM provides optimal performances using the whole dataset. These changing characteristics emphasize that optimal predictions can only be obtained using dynamic algorithm selection, given the nature of the data at hand.

The integration of MLOps practices via MLflow facilitated efficient model versioning, tracking, and deployment, demonstrating the framework’s adaptability to evolving data patterns. This adaptability is crucial for real-world applications, where seismic activity is subject to temporal and spatial variations. The ability to dynamically select the best-performing algorithm ensures that the pipeline remains effective as data availability and quality change over time.

By using different kinds of 3D visualizations, feature correlation heatmaps, and the distribution depth, analysis has illustrated useful relationships between earthquake occurrences over the years. These visualizations help highlight the non-linear nature of these interactions in seismic data and architecture selection and guide our feature-level choices for a predictive model. The correlation between depth and magnitude, as well as the annual trends in the number of earthquakes, also offer important context to bolster the interpretation and utilization of our predictive models.

This study underscores the effectiveness of machine learning, particularly gradient-boosting techniques, in predicting natural calamities. The deployment of these models within an MLOps framework ensures practical applicability and long-term sustainability in dynamic environments. In future work, we plan to integrate advanced MLOps tools, such as automated feature engineering and hyper-parameter optimization frameworks, to enhance model performance further. Additionally, incorporating real-time data pipelines into our MLOps framework will enable faster adaptation to new data, ensuring that the models remain accurate and relevant in rapidly evolving conditions.