Multi-Task Learning Convolutional Recurrent Neural Network (MTL-CRNN) for Joint Estimation of Geomagnetic Earthquake Precursor Parameters

Accurate earthquake precursor identification remains a formidable challenge in geophysics due to the non-linear complexity of geomagnetic signals and the high risk of data leakage in predictive modeling. Earthquakes cause massive destruction due to their extreme magnitude and intensity, making the prediction of precursors—such as electromagnetic (EM) field alterations—essential for EarlyWarning Systems (EWS). This paper proposes a novel Multi-Task Learning Convolutional Recurrent Neural Network (MTL-CRNN) framework utilizing a headless EfficientNet-B1 coupled with a Bidirectional Gated Recurrent Unit (BiGRU) for the joint estimation of earthquake parameters in Indonesia. To preserve physical signal integrity, three-component geomagnetic data (H, D, and Z) are transformed into raw spatio-temporal Continuous Wavelet Transform (CWT) tensors and compiled into a monolithic FP16 HDF5 structure. Unlike conventional single-task approaches, our MTL-CRNN architecture simultaneously optimizes three predictive heads: binary precursor detection, magnitude classification, and circular azimuth regression. A rigorous chronological data splitting strategy with a 30-day embargo period was implemented to eliminate forward-looking data leakage. To combat space-weather interference, a novel Geomagnetic Penalty Loss (GPL) was introduced; empirical results demonstrate that GPL significantly suppresses non-tectonic False Positives during intense solar storms, enabling the model to achieve an operational precision of 61.5% while maintaining a 100.0% Recall (Zero False Negatives) for impending seismic events.A massive national-scale blind audit across 19 BMKG stations (comprising 1,623 daily samples with 28 exclusive 2026 blind-test cases) demonstrated that the model achieved a remarkable Sensitivity (Recall) of 100.0% for precursor detection. While maintaining high sensitivity, the proposed Geomagnetic Penalty Loss (GPL) significantly suppressed False Positives during documented solar storms, achieving an operational precision of 61.5% and a Mean Absolute Angular Error (MAAE) of ±72.32° for azimuth estimation.