Earthquakes represent one of the most consequential natural hazards facing human civilisation, causing catastrophic loss of life, infrastructure destruction, and long-lasting economic disruption with little or no advance warning. Existing seismic monitoring infrastructure excels at real-time event detection but offers minimal near-term forecasting capability, while academic machine learning research has largely failed to translate into accessible, deployable tools for the emergency management community. This paper presents SeismoPredict AI, an intelligent seismic activity forecasting system that addresses these gaps by integrating two complementary predictive methodologies -Long Short-Term Memory (LSTM) recurrent neural networks and ARIMA (Autoregressive Integrated Moving Average) statistical timeseries modelswithin a unified, interactive Streamlit web application. The system architecture implements a six-module processing pipeline: a Data Collection Module for CSV-format earthquake catalog ingestion using Pandas; a Preprocessing Module for missing value removal and chronological sorting; a Feature Engineering Module applying a sliding-window series-to-supervised conversion (lookback n=3) to generate LSTM training arrays; a Model Training Module implementing a two-layer Keras Sequential LSTM (64 units, Adam optimiser, MSE loss, 10 epochs) and an ARIMA(5,1,0) model via statsmodels, with Joblib/HDF5 persistence; a Prediction Module with automatic model retraining fallback when persisted files are unavailable; and an Alert Module providing SMTP-based email notification when forecasted magnitude exceeds configurable risk thresholds. Empirical evaluation on USGS-sourced earthquake catalogs spanning three seismically active regions demonstrates that the LSTM model achieves a Root Mean Squared Error (RMSE) of 0.312 magnitude units and a Mean AbsoluteError (MAE) of 0.241 magnitude units on held-out test sequences, outperforming the ARIMA(5,1,0) baseline (RMSE = 0.389, MAE = 0.298) on non-linear magnitude patterns while the ARIMA model provides superior interpretability and 6.4 faster inference.The Streamlit dashboard renders complete end-to-end workflowsfrom data upload through interactive Plotly visualisation to five-step ARIMA forecast displaywithin 18.3 seconds mean completion time on consumer hardware, with user acceptance testing confirming a mean usability rating of 4.2 out of 5.0 across six assessment dimensions..
Jayasree et al. (Thu,) studied this question.