BIBECHANA Vol. 22, No. 1, April 2025, 22-29 ISSN 2091-0762 (Print), 2382-5340 (Online) Journal homepage: http://nepjol.info/index.php/BIBECHANA Publisher:Dept. of Phys., Mahendra Morang A. M. Campus (Tribhuvan University)Biratnagar Machine learning for predicting earthquake magnitudes in the Central Himalaya Ram Krishna Tiwari1, Rudra Prasad Poudel1,2, Harihar Paudyal1 1Birendra Multiple Campus, Tribhuvan University, Bharatpur, Chitwan, Nepal 2Central Department of Physics, Tribhuvan University, Kirtipur, Kathmandu, Nepal ∗Corresponding author. Email: ram.tiwari@bimc.tu.edu.np Abstract Human intervention cannot halt natural disasters like earthquakes, but machine learning ap- plications expertise can be utilized to detect patterns in data and increase understanding and predictive power. Recent development of machine learning models has increasingly developed interest in forecasting and predicting the magnitude of earthquakes. In this work, Random Forest Regressor (RFR), Multi-Layer Perceptron Regressor (MLPR), and Support Vector Regression (SVR) models were employed to predict the magnitude of greater than 6 mb earth- quakes that occurred in the year 2015 in the central Himalaya. We noticed RFR method had been able to predict the magnitude of the Gorkha earthquake (6.9 mb), the Kodari earthquake (6.7 mb), and 6.5 mb magnitude earthquake (aftershock of Gorkha earthquake) in comparison with the other two models. We also checked the performance of these models by three parame- ters Mean Absolute Error (MAE), Mean Squared Error (MSE), and Root Mean Squared Error (RMSE) and noticed the better performance of RFR model. The findings illustrate that RFR is achieving better performance than the other two algorithms, as the predicted magnitudes are close to the actual magnitudes. Keywords Machine Learning, Earthquake, Regressor, Prediction. Article information Manuscript received: October 6, 2024; Revised: January 7, 2025; Accepted: January 12, 2025 DOI https://doi.org/10.3126/bibechana.v22i1.70637 This work is licensed under the Creative Commons CC BY-NC License. https://creativecommons. org/licenses/by-nc/4.0/ 1 Introduction An earthquake is a natural disaster that strikes suddenly, between seconds to minutes, and shakes a large area of landmass, potentially killing peo- ple and damaging property. Nepal, which is posi- tioned in the center of the Himalayan arc, saw many small and large earthquakes in last millennia [1–5]. The seismic activity in the Himalayan region is im- pacted by the buildup of strain energy that hap- pened roughly 50 million years ago during the In- dian plate's thrust beneath the Eurasian plate [6–8]. 22 http://nepjol.info/index.php/BIBECHANA ram.tiwari@bimc.tu.edu.np https://doi.org/10.3126/bibechana.v22i1.70637 https://creativecommons.org/licenses/by-nc/4.0/ https://creativecommons.org/licenses/by-nc/4.0/ Ram Krishna Tiwari et al./ BIBECHANA 22 (2025) 22-29 23 The region has had more recent earthquakes in the past 70 years, including the 1988 Udayapur earth- quake of magnitude 6.6 Mw, 2011 Sikkim earth- quake of magnitude 6.9 Mw, 2015 Gorkha earth- quake of magnitude 7.9 Mw, Dolakha (Kodari) earthquake of magnitude 7.3 Mw, Doti earthquake of magnitude 6.6 ML, and the 2023 Jajarkot earth- quake of magnitude 6.4 ML [8–10]. Human intervention cannot halt natural disas- ters like earthquakes, but machine learning appli- cation expertise can be utilized to detect patterns in data and increase understanding and predictive power [11–13]. Most of the machine learning (ML) algorithms fall into one of two categories: Super- vised Learning (SL) and Unsupervised Learning (USL) (Figure 1). Figure 1: Basic idea of selecting the ML algorithms. Unsupervised ML for unlabelled data, and super- vised ML for labelled data. In SL, the computer is instructed or trained with the labeled data. SL algorithms construct two types of predictive models, Regression and Classification models which approaches data in a different way. For forecasting a numerical value, regression model is used. USL includes utilizing an unlabeled dataset to train the computer, after which it uses its own judgment to anticipate outputs. The primary goal of unsupervised learning is to categorize or group the unsorted dataset according to similarities, dif- ferences, and patterns. The machines are to find the hidden patterns in the input dataset [14]. In contrast to traditional methodologies, ma- chine learning (ML) models offer a fresh and cre- ative way to find hidden signals and patterns. This development covers a wide range of seismic appli- cations, such as earthquake detection and phase identification, early warning systems, ground mo- tion prediction, seismic tomography, earthquake geodesy, seismic risk assessment, and finally earth- quake prediction [15,16]. It is generally accepted that there isn't a sin- gle ideal algorithm or machine learning solution that works for all situations and datasets because algorithm performance varies on a variety of pa- rameters. While some algorithms work better with tiny amounts of data, others are more effective with large samples of data. While some algorithms just need quantitative inputs, others demand categori- cal inputs. The complication of the data and the number of features that the model needs to un- derstand and make predictions are crucial factors when picking an algorithm. To account for this, three distinct algorithms namely, Random Forest Regressor (RFR), Support Vector Regressor (SVR), Multi-Layer Perceptron Regressor (MLPR) have been used in this work to analyse an earthquake dataset [17]. Different hyper parameters have been looked at for each model that has been selected, and the predicted results have been equitably as- sessed with metrics like Mean Square Error (MSE), Root Mean Square (RMSE and Mean Absolute Er- ror (MAE). 1.1 Random Forest A random forest algorithm builds a forest from many separate decision trees, and chooses the out- come based on the predictions offered by them. The decision nodes, leaf nodes, and root nodes make up a decision tree (Figure 2). Figure 2: Schematic presentation of Random Forest algorithm. Decision trees are trained using knowledge ac- quisition. A method splits a training dataset into branches, and then further divides those branches until a leaf node is reached [18]. Splitting branches during the creation of decision trees depends on en- tropy and information gain. The attributes used to forecast the outcome are represented by the nodes in the decision tree. Ram Krishna Tiwari et al./ BIBECHANA 22 (2025) 22-29 24 1.2 Support Vector Regression A Support Vector Regressor (SVR) is a mathemat- ical framework designed to function as a method or strategy for optimizing a particular mathemat- ical function concerning a provided dataset. The SVR method focuses on finding the ideal hyper- plane in the feature space (N-dimensional space) which in turns divides the data points into various classes [19]. The hyperplane seeks to make the dis- tance between the closest points of various classes as large as feasible. The size of the hyperplane is de- termined by the quantity of features. For two input characteristics, the hyperplane essentially becomes a line whereas it turns into a 2-D plane for three input characteristics. 1.3 Multi-Layer Perceptron Regression A Multi-Layer Perceptron regression (MLPR) con- sists of at least three layers: an input layer, a hid- den layer, and an output layer. Every layer makes use of the results from the layer before it. Without input, each layer node is referred to as a neuron. The primary processing unit of the neural network, the neuron, gathers data from a variety of inputs, applies weights and bias terms, and then sends the final product to an active function that produces outputs [20]. A multilayer perceptron model pri- marily comprises of a back propagation model for training, and additionally, it employs a linear ac- tivation function in its hidden layers, which, when combined with multiple layers, allows it to approx- imate non-linear relationships in the data. Figure 3: Schematic presentation of Multilayer Per- ceptron algorithm where L0 stands for (Layer 0) or Input Layer (IL), L1 stands for (Layer 1) or Hidden Layer (HL), and L2 stands for (Layer 2) or Output Layer (OL). The study area is in the central Himalaya region between latitudes of 26.5° and 30.5° and longitudes of 80° and 88°, which includes the entirety of Nepal along with certain regions of India and China. The region has low to moderate seismicity (Figure 4). The seismicity of the region is primarily controlled by the Main Central Thrust (MCT), Main Bound- ary Thrust (MBT), Main Frontal Thrust (MFT), and several small faults that are trending north to the south [21–23]. Figure 4: Seismicity of the study area. Also show- ing major Himalayan thrusts STD, MCT, MBT, and MFT from north to south, namely South Ti- betan Detachment System, Main Central Thrust, Main Boundary Thrust, Main Frontal Thrust. The earthquake occurrence mechanisms are widely regarded as unpredictable and marked by non-linear behaviors. Most ongoing research in the field of ML has concentrated on exploring how neu- ral networks can be applied to address this challenge [11,24,25]. Because of conceptual, algorithmic, and computational constraints, it was challenging to construct efficient models via the early explorations, but it is now possible to use advanced models by leveraging Deep Neural Networks (DNNs) [13,26]. While going through the literature, Artificial Neural Network (ANN) in collaboration with seis- mic precursors were found useful to predict earth- quakes. For example, Back Propagation Neural Network (BPNN) models have been used to iden- tify unusual behaviour in radon concentrations pro- duced by earthquakes [27]. From the assembly of Radial Basis function (RBF) neural networks, earthquakes in China have been predicted (Y. Liu et al., 2004). Analysing seismic events over a period in southern California and San Francisco, a model was designed, based on ANNs which can predict the earthquakes on monthly basis [16]. In 2009, the identical seismic considerations were incorporated with the Probabilistic Neural Network (PNN) to predict earthquake [15]. The work [28] demonstrates that, although the occurrence of earthquakes is nonlinear and ran- dom phenomena, it is still possible to model it us- ing methodologies of machine learning. The neural network-based method for predicting earthquakes was evaluated using data from the Portuguese re- Ram Krishna Tiwari et al./ BIBECHANA 22 (2025) 22-29 25 gion of the Azores, and the results showed that it successfully predicted earthquakes in July 1998 for the modified Mercalli intensity (MMI) of 8, and in January 2004 for the modified Mercalli intensity (MMI) of 5, respectively [29]. According to stud- ies [15, 16] probabilistic neural networks (PNNs) may be used for small and intermediate earthquake prediction while recurrent neural networks (RNNs) may be utilized for earthquakes of large magni- tude. According to a comparative study employ- ing ANN technology and non-linear predictability assessments, the dynamics of earthquakes in the North-East India region were found to be stochas- tically scaled process [30]. In a study conducted in Hindukush region by the application of four ma- chine learning algorithms on a temporal distribu- tion of past earthquakes, the accuracy to predict the earthquakes is noticed to be 65% on Linear Pro- gramming Boost Ensemble while 58% and 62% for RNN and Random Forest [31]. The aim of this study is to utilize historical seis- mic data, including date, time, latitude, and longi- tude to create predictive algorithms that can pre- dict the magnitude and compare it with the actual magnitude of the earthquakes in central Himalayan region. The primary challenge is to design and train machine learning models that enhances our ability to predict earthquake magnitudes accurately, aid- ing in disaster preparedness and response efforts. 2 Data and Methodology This research is quantitative and based on quan- titative earthquake data like magnitude, latitude, longitude, and focal depth. The data extracted from the International Seismological Centre (ISC) catalog for the period between February 1, 1964, and December 27, 2022 [32], included additional entries such as event identification numbers, au- thor names, station codes, and phase data, which were deemed irrelevant for this study and ex- cluded. The catalog was further processed to iden- tify missing values in critical attributes like magni- tude, depth, latitude, and longitude, with incom- plete records removed. The final dataset comprises 2595 earthquakes, with magnitudes ranging from 2.9 to 6.9. After compilation, the data was thor- oughly cleaned, formatted, and stripped of miss- ing or damaged entries to ensure its suitability for analysis. The three models, namely Random Forest Regressor (RFR), Multi-Layer Perceptron Regres- sor (MLPR), and Support Vector Regressor (SVR) are selected over others for their proven capability to manage the inherent complexities of the dataset, such as non-linearity, high-dimensionality, and data noise [33, 34]. The data processing in this study involves configuring key hyper parameters across these models. For RFR, bootstrapping is enabled (bootstrap=True), and the model uses all features for each split (max_features=1.0), with trees grow- ing until leaves are pure or contain fewer than 2 samples (min_samples_split=2). It employs the 'squared_error' criterion to minimize mean squared error and uses 100 trees (n_estimators=100). The MLPR is configured with the 'tanh' ac- tivation function, a learning rate of 0.001 (learn- ing_rate_init=0.001), and 500 neurons in the hid- den layer (hidden_layer_sizes=500), utilizing the 'sgd' solver for optimization. It runs without early stopping (early_stopping=False) and continues un- til reaching the maximum number of iterations (max_iter=100). The SVR model uses an RBF kernel (kernel='rbf') with a regularization param- eter (C=1.0) and an epsilon of 0.1 to control er- ror margin, while gamma is set to 'auto'. All models are trained with random_state=0 for re- producibility, and verbose output is suppressed (verbose=False). These configurations collectively guide the data through preprocessing, transforma- tion, and optimization steps to ensure that the ma- chine learning models are effectively trained and evaluated.80% of the dataset was provided for train- ing and the remaining 20% for testing, allowing for proper model evaluation on unseen data (Aryal et al., 2024). Thereafter, numerical optimization al- gorithms are employed to iteratively fine-tune the model parameters using a cost function. To eval- uate the model's performance, we calculate met- rics such as Mean Absolute Error (MAE), Mean Square Error (MSE), and Root Mean Square Er- ror (RMSE). Finally, the machine learning model is employed to handle new data. 3 Results and Discussion Machine learning offers promising capabilities for analyzing historical data to predict earthquake magnitudes, a task with no definitive method yet. This research leverages data from 2595 seis- mic events and applies Random Forest Regres- sor (RFR), Support Vector Regressor (SVR), and Multi-Layer Perceptron Regressor (MLPR). These models were chosen for their ability to handle non- linear relationships and complex geophysical pat- terns: RFR for robustness and feature analysis, SVR for modeling nonlinearities with kernel meth- ods, and MLPR for capturing intricate dependen- cies. Their selection is supported by their proven reliability in similar seismological studies [33–35]. We have tested these methods to forecast the mag- nitude of the earthquakes that hit the central Hi- malaya region in the year 2015 and comparison be- tween actual magnitude and predicted magnitude are depicted by the different plots. The density magnitude plot of the dataset is depicted by Figure 5. Ram Krishna Tiwari et al./ BIBECHANA 22 (2025) 22-29 26 Figure 5: A density magnitude plot provides a vi- sual representation of the frequency of earthquakes at different magnitudes. The Y-axis represents the density of earth- quakes, often measured as the number of earth- quakes per unit magnitude. This shows how fre- quently earthquakes of different magnitudes occur within the dataset. It tends to have a higher den- sity for smaller magnitude earthquakes and gradu- ally decreases as the magnitude increases. A sin- gle higher peak in the plot suggests that 3.1 to 5.4 magnitudes are predominant in that region. The plot shows a steep decrease in density as magni- tude increases which suggests that there are fewer large earthquakes, a typical characteristic of most seismically active regions. Five earthquakes of the year 2015, namely the Gorkha earthquake (6.9 mb), the Kodari earth- quake (6.7 mb), the magnitude 6.1 mb earthquake, the magnitude 6.5 mb earthquake and the magni- tude 6.6 mb earthquake of diverse locations, are used to assess the performance of the proposed tech- niques. The RFR, MLPR, SVR models are trained and tested through multiple cycles to refine their performance which helps to get the minimum er- rors [33,34,36,37]. Table 1 gives the predicted mag- nitude and actual magnitude of above-mentioned earthquakes and heat map of predicted magnitude is presented in Figure 6. Figure 6: A heat map of magnitude predicted by Random Forest (RFR), Multi-Layer Perceptron (MLPR), Support Vector Regression (SVR) for the Gorkha earthquake, the Kodari earthquake, the magnitude 6.1 earthquake, the magnitude 6.5 earth- quake and the magnitude 6.6 earthquake. Following the training phase, the model is val- idated to forecast the earthquake exceeding 6.0 in mb scale and the 3d plot of the predicted magnitude and actual magnitude plot are depicted by Figure 7. Table 1: Predicted and actual magnitude of the earthquakes. Name RFR based magnitude MLPR based magnitude SVR based magnitude Actual magnitude Gorkha EQ 6.24 4.09694 4.61592 6.9 Kodari EQ 5.846 4.08975 4.32297 6.7 EQ6.1 3.732 4.06169 3.86912 6.1 EQ6.5 5.887 4.10614 4.47054 6.5 EQ6.6 4.629 4.09994 4.18434 6.6 Table 2: Adapted machine learning model with error and accuracy. Model MAE MSE RMSE Accuracy of Model Random Forest Regressor (RFR) 0.36 0.23 0.48 0.16 Multi-layer Perceptron Regressor (MLPR) 0.40 0.27 0.52 0.016 Support Vector Regressor (SVR) 0.35 0.24 0.49 0.13 Ram Krishna Tiwari et al./ BIBECHANA 22 (2025) 22-29 27 The magnitude predicted by RFR is close to the actual magnitude of the Gorkha earthquake, the Kodari earthquake, and the magnitude 6.5 after- shock (above 6.0 for both events) while the magni- tude predicted for other earthquakes deviates from the actual value (Figure 7). The magnitude pre- dicted by SVR significantly deviates from the ac- tual values, as it predicts 4.61592 for the Gorkha earthquake and 4.32297 for the Kodari earthquake. The magnitude predicted by MLPR is just around 4 and almost the same for all five earthquakes and greatly deviated from the actual magnitudes. Figure 7: 3D plot of predicted magnitude versus actual magnitude for the Gorkha earthquake, the Kodari earthquake, the magnitude 6.1 earthquake, the magnitude 6.5 earthquake, and the magnitude 6.6 earthquake using RFR, MLPR, SVR. 3.1 Performance Evaluation The performance of RFR, SVR, and MLPR is pre- sented in Table 2 and Figure 8 and their accuracies are presented in Figure 9. The 3D error bar in Figure 8 depicts the errors produced by RFR, MLPR, and SVR models. The MAE, MSE, and RMSE errors for RFR (0.36, 0.23, 0.48) and SVR (0.35, 0.24, 0.49) exhibit minimal disparity, whereas MLPR's errors are marginally el- evated (0.40, 0.27, 0.52). These findings indicate that MLPR's performance is comparatively inferior to that of RFR and SVR. The RFR model demonstrates superior perfor- mance in predicting earthquake magnitudes com- pared to the other two methods. The accuracy plot reveals that RFR achieves a slightly higher accuracy (0.16) than SVR (0.13) and MLPR (0.016) (Figure 9). This can be attributed to RFR's ensemble ap- proach, which leverages decision trees to capture complex patterns in the data effectively. In con- trast, SVR prioritizes a smoother fit by minimiz- ing margin violations, which may limit its ability to capture intricate details. The notably low accu- racy of the MLPR model suggests that it fails to learn meaningful patterns from the data, leading to predictions that are essentially random. Figure 8: 3D error bar of three different methods employed for the estimation of the magnitudes. Figure 9: Accuracy bar of three different methods used for the estimation of magnitudes. 4 Conclusion Three machines learning algorithms namely, Ran- dom Forest Regressor (RFR), Multi-layer Percep- tron (MLPR), and Support Vector Regressor (SVR) have been used to predict the major earthquake events that occurred in the year 2015. The study has been applied to 2595 earthquakes of magnitude range 2.9 to 6.9, collected from ISC catalog, for 68 years. The results suggest that all algorithms, while providing reasonable estimates, tend to un- der predict earthquake magnitudes, particularly for higher-magnitude events. For example, the RFR model predicts 6.24 for Gorkha earthquake (actual 6.9) and 5.85 for Kodari earthquake (actual 6.7), indicating that it tends to slightly underestimate magnitudes. Similarly, the MLPR model under pre- dicts with values like 4.10 for Gorkha earthquake, while the SVR model also provides lower predic- tions, such as 4.62 for Gorkha earthquake. Among three algorithms RFR is found to be the superior as it estimates the magnitude close to actual magni- tude. This highlights the need for further model re- finement through hyper parameter tuning, feature engineering, or incorporating more detailed seis- mic data to enhance prediction accuracy, especially for larger earthquakes. Analyzing the residuals be- tween predicted and actual values could help iden- Ram Krishna Tiwari et al./ BIBECHANA 22 (2025) 22-29 28 tify areas where the models are underperforming and guide improvements. From the error perspective both RFR and SVR do not show significant differences but as the pre- diction pattern is observed RFR shows the better promises. Despite challenges such as imbalanced datasets, uncertainties in seismic features, and the complexity of modeling dynamic, non-linear rela- tionships in earthquake patterns, the results of this study indicate that both the Random For- est Regressor (RFR) and Support Vector Regres- sor (SVR) hold promise for prediction, particularly when large datasets are available. Acknowledgments One of the authors (RKT) would like to acknowl- edge the Research Management Cell (RMC), Biren- dra Multiple Campus, Tribhuvan University, Nepal for providing financial support. References [1] R. V. Almeida, J. Hubbard, L. Liberty, A. Fos- ter, and S. N. Sapkota. Seismic imaging of the main frontal thrust in nepal reveals a shallow décollement and blind thrusting. Earth and Planetary Science Letters, 494:216–225, 2018. [2] M. Hubbard, M. Mukul, A. P. Gajurel, A. Ghosh, V. Srivastava, B. Giri, N. Seifert, and M. M. Mendoza. Orogenic segmentation and its role in himalayan mountain building. Frontiers in Earth Science, 9, 2021. [3] K. M. Khattri and A. K. Tyagi. Seismic- ity patterns in the himalayan plate boundary and identification of the areas of high seis- mic potential. Tectonophysics, 96(3–4):281– 297, 1983. [4] M. Nayak and T. G. Sitharam. Estimation and spatial mapping of seismicity parameters in western himalaya, central himalaya and indo- gangetic plain. Journal of Earth System Sci- ence, 128(3), 2019. [5] R. K. Tiwari. Multifractal Approach To the Study of Gorkha Earthquake of 25 April, 2015 Nepal. PhD thesis, Institute of Science Tech- nology, Tribhuvan University, 2023. [6] A. K. Dubey. Understanding an Orogenic Belt. Springer, 2014. [7] J. F. Ni. Active tectonics of the himalaya. Proceedings of the Indian Academy of Sciences - Earth and Planetary Sciences, 98(1):71–89, 1989. [8] R. K. Tiwari and H. Paudyal. Geodynamics of gorkha earthquake (mw 7.9) and its after- shocks. Himalayan Physics, 9:103–109, 2020. [9] G. Monsalve, A. Sheehan, V. Schulte-Pelkum, S. Rajaure, M. R. Pandey, and F. Wu. Seis- micity and one-dimensional velocity structure of the himalayan collision zone: Earthquakes in the crust and upper mantle. Journal of Geo- physical Research: Solid Earth, 111(10):1–19, 2006. [10] R. K. Tiwari and H. Paudyal. Spatial map- ping of b-value and fractal dimension prior to november 8 , 2022 doti earthquake, nepal. PLOS ONE, 18(8):1–13, 2023. [11] M. H. Al Banna, K. A. Taher, M. S. Kaiser, M. Mahmud, M. S. Rahman, A. S. M. S. Ho- sen, and G. H. Cho. Application of artificial intelligence in predicting earthquakes: State- of-the-art and future challenges. IEEE Access, 8:192880–192923, 2020. [12] M. Aryal, R. K. Tiwari, and H. Paudyal. Pre- diction of earthquakes in nepal and the adjoin- ing regions using lstm. BMC Journal of Scien- tific Research, 7:12–26, 2024. [13] P. M. R. DeVries, F. Viégas, M. Wattenberg, and B. J. Meade. Deep learning of aftershock patterns following large earthquakes. Nature, 560(7720):632–634, 2018. [14] S. Shalev-Shwartz and S. Ben-David. Under- standing machine learning: From theory to al- gorithms. Cambridge University Press, 2014. [15] H. Adeli and A. Panakkat. A probabilistic neu- ral network for earthquake magnitude predic- tion. Neural Networks, 22(7):1018–1024, 2009. [16] A. Panakkat and H. Adeli. Neural network models for earthquake magnitude prediction using multiple seismicity indicators. Interna- tional Journal of Neural Systems, 17(1):13–33, 2007. [17] G. Asencio-Cortés, F. Martínez-Álvarez, A. Troncoso, and A. Morales-Esteban. Medium–large earthquake magnitude pre- diction in tokyo with artificial neural net- works. Neural Computing and Applications, 28(5):1043–1055, 2017. [18] K. Budiman and Y. N. Ifriza. Analysis of earthquake forecasting using random forest. Journal of Soft Computing Exploration, 2(2), 2021. [19] J. Tezcan and Q. Cheng. Support vector re- gression for estimating earthquake response spectra. Bulletin of Earthquake Engineering, 10(4):1205–1219, 2012. Ram Krishna Tiwari et al./ BIBECHANA 22 (2025) 22-29 29 [20] J. Mahmoudi, M. A. Arjomand, M. Rezaei, and M. H. Mohammadi. Predicting the earth- quake magnitude using the multilayer percep- tron neural network with two hidden layers. Civil Engineering Journal, 2(1):1–12, 2016. [21] J. Liu, C. Ji, J. Zhang, P. Zhang, L. Zeng, Z. Li, and W. Wang. Tectonic setting and general features of coseismic rupture of the 25 april, 2015 mw 7.8 gorkha, nepal earth- quake. Chinese Science Bulletin, 60(27):2640– 2658, 2015. [22] R. K. Tiwari, H. Paudyal, and D. Shanker. On the spatio-temporal variation in b-value after 25 april 2015 gorkha, nepal earth- quake. Geodesy and Geodynamics, 13(5):525– 533, 2022. [23] B. N. Upreti. An overview of the stratigraphy and tectonics of the nepal himalaya. Journal of Asian Earth Sciences, 17(5–6):577–606, 1999. [24] G. C. Beroza, M. Segou, and S. Mostafa Mousavi. Machine learning and earthquake forecasting—next steps. Nature Communications, 12(1):10–12, 2021. [25] S. M. Mousavi and G. C. Beroza. Machine learning in earthquake seismology. Annual Re- view of Earth and Planetary Sciences, 51:105– 129, 2023. [26] T. Chelidze, G. Melikadze, T. Kiria, T. Jimshe- ladze, and G. Kobzev. Statistical and non- linear dynamics methods of earthquake fore- cast: Application in the caucasus. Frontiers in Earth Science, 8(June), 2020. [27] A. Negarestani, S. Setayeshi, M. Ghannadi- Maragheh, and B. Akashe. Layered neural net- works based analysis of radon concentration and environmental parameters in earthquake prediction. Journal of Environmental Radioac- tivity, 62(3):225–233, 2002. [28] K. M. Asim, F. Martínez-Álvarez, A. Basit, and T. Iqbal. Earthquake magnitude predic- tion in hindukush region using machine learn- ing techniques. Natural Hazards, 85(1):471– 486, 2017. [29] E. I. Alves. Earthquake forecasting using neu- ral networks: Results and future work. Non- linear Dynamics, 44(1-4):341–349, 2006. [30] S. Sri Lakshmi and R. K. Tiwari. Model dissec- tion from earthquake time series: A compara- tive analysis using modern non-linear forecast- ing and artificial neural network approaches. Computers and Geosciences, 35(2):191–204, 2009. [31] R. Mallouhy, C. A. Jaoude, C. Guyeux, and A. Makhoul. Major earthquake event pre- diction using various machine learning al- gorithms. In 6th International Conference on Information and Communication Technolo- gies for Disaster Management, ICT-DM 2019, 2019. [32] D. Di Giacomo, E. Robert Engdahl, and D. A. Storchak. The isc-gem earthquake cat- alogue (1904-2014): Status after the exten- sion project. Earth System Science Data, 10(4):1877–1899, 2018. [33] R. Jain, A. Nayyar, S. Arora, and A. Gupta. A comprehensive analysis and prediction of earthquake magnitude based on position and depth parameters using machine and deep learning models. Multimedia Tools and Appli- cations, 80(18):28419–28438, 2021. [34] C. N. Schuba, J. P. Schuba, G. G. Gray, and R. G. Davy. Interface-targeted seismic velocity estimation using machine learning. Geophysi- cal Journal International, 218(1):45–56, 2019. [35] T. Y. Hsu and A. Pratomo. Early peak ground acceleration prediction for on-site earthquake early warning using lstm neural network. Fron- tiers in Earth Science, 10(July):1–17, 2022. [36] E. Harirchian and T. Lahmer. Improved rapid assessment of earthquake hazard safety of structures via artificial neural networks. In IOP Conference Series: Materials Science and Engineering, volume 897, 2020. [37] B. Nadi, F. Askari, O. Farzaneh, S. Fato- lahzadeh, and R. Mehdizadeh. Reliability eval- uation of regression model for estimating co- seismic landslide displacement. Iranian Jour- nal of Science and Technology - Transactions of Civil Engineering, 44(1):165–173, 2020. Introduction Random Forest Support Vector Regression Multi-Layer Perceptron Regression Data and Methodology Results and Discussion Performance Evaluation Conclusion