Data Driven Earthquake Prediction using Intelligent Techniques

Abstract

Earthquakes represent one of the most devastating natural hazards, claiming thousands of lives and causing trillions in economic damage annually, particularly in seismically active regions like the Himalayan arc and Pacific Ring of Fire . Despite advances in seismic monitoring, accurate prediction of earthquake magnitude the key determinant of potential destruction remains elusive due to the nonlinear, chaotic dynamics of fault systems and the rarity of large events in historical catalogs. In conventional probabilistic seismic hazard analysis, long term risk predictions can be made, but reliable short term predictions are not possible, which makes society prone to abrupt ruptures. This thesis works to fill this crucial research gap by proposing and developing a new deep learning model to predict magnitude based on multi modal precursor. Although there have been some recent attempts to apply machine learning to seismicity forecasting, there have been several models still using simple regression models or simple neural networks, which are not capable of handling spatiotemporal correlations effectively. Although models such as random forest approaches have had moderate success at the regional scale, they also tend to neglect long term temporal correlations and lack generability across different tectonic settings. Although there have been some promising approaches using transformers for time series data, there have been no attempts to apply these models to seismicity forecasting, mainly due to their high complexity and the need for specific adaptations. These models, although successful, also emphasize the need to develop models that can integrate heterogeneous features of seismicity, such as b value anomaly, ionospheric TEC, and waveform magnitude, effectively, which can also provide probabilistic forecasting. In light of these challenges, this thesis develops the approach for earthquake magnitude prediction using modern deep learning and making a particular focus on: revealing temporal dynamics and predictive uncertainty. It focuses on an enhanced Temporal Fusion Transformer (TFT) architecture combining bidirectional LSTMs, GRUs, multi head attention, and gating mechanisms that combine past observations with simple future covariates and static spatial contextual information into a single unified forecasting model. To overcome the small size and imbalance in conventional seismic catalogs, this study constructs an augmented dataset with a model based synthetic data generator and relies on the assumption that the added records preserve the statistical and spatial properties of the original events. The original data and augmented are subjected to extensive feature engineering comprising lagged magnitudes, rolling statistics, interv action terms, and cluster based spatial descriptors, followed by their transformation into supervised sequences to model time series. The TFT model is trained using a quantile loss function to produce probabilistic forecasts at multiple quantiles, allowing the derivation of both point estimates and prediction intervals for earthquake magnitude. The training procedure incorporates regularization techniques such as dropout, weight decay, and gradient clipping, together with adaptive learning rate scheduling and validation based check pointing to promote stable convergence and generalization. Model performance is evaluated on a temporally held out test set using root mean squared error (RMSE), mean absolute error (MAE), and average quantile loss, alongside visual diagnostics such as prediction truth scatter plots, residual analysis, and time series overlays. In addition, Monte Carlo dropout is employed at inference time to form an ensemble of stochastic predictions, providing an estimate of epistemic uncertainty and enabling a comparison between single model and ensemble behavior. The results indicate that the proposed TFT architecture, combined with carefully designed augmentation and feature engineering, can achieve low prediction errors while generating well calibrated prediction intervals for normalized earthquake magnitudes. Spatial comparison plots show that the synthetic events closely follow the geographic distribution of the original Catalog, suggesting that the augmented dataset is suitable for training without distorting the underlying seismic patterns. Overall, the thesis demonstrates that attention based sequence models, when paired with realistic data augmentation and rigorous evaluation, offer a promising direction for probabilistic earthquake magnitude forecasting and for quantifying the uncertainty inherent in such predictions.