Deep Learning-Based Probabilistic Earthquake Forecasting Along the Tan-Lu Fault Zone Using Multiparameter Seismicity Features

Here, we develop a supervised framework for identifying probabilistic earthquake-risk states in the central and southern Tan-Lu fault zone and adjacent areas of eastern China using multiple catalog-derived seismicity features and a deep learning approach. Using the China Earthquake Networks Center catalog from 1979 to 2025, seven seismological parameters (b, M, η, D, Mf, N, and Rm) were calculated for ML≥2.0 earthquakes. These parameters were transformed into probability-based time-series features and used as inputs to an enhanced long short-term memory model with dual-task learning and dropout regularization. The model was trained using data from 1979 to 2020 and evaluated on an independent testing period from 2021 to 2025. Training was repeated using five predefined random seeds, and same-date predicted probabilities were averaged with equal weights. In the broad study area, the seven-parameter probability-transformed ensemble achieved an area under the receiver operating characteristic curve (ROC-AUC) of 0.650 and average precision (AP) = 0.297, compared with 0.535 and 0.258 for the raw-parameter ensemble; the no-skill AP baseline was 0.209. Input-feature ablation showed that the full seven-parameter model outperformed models using N alone or N together with b. Comparison with an epidemic-type aftershock sequence model indicated that the relative performance of statistical and data-driven models depends strongly on the forecasting task. Although elevated scores were not restricted to labeled pre-earthquake windows and the separation between positive and negative score distributions weakened in the independent test period, the continuous scores retained some ranking information. Accordingly, the model output is interpreted as a continuous risk-ranking score rather than as a deterministic forecast of earthquake occurrence on any specific date. Probability-based representations of multiple catalog-derived seismicity parameters, combined with deep temporal modeling, may provide a useful complementary approach to regional identification of pre-earthquake risk windows.