This study presents a systematic evaluation of four neural network architectures for aftershock location forecasting, ranging from a shallow single-hidden-layer model to a deep six-hidden-layer network, using identical input features to isolate the effect of architectural depth. The models were rigorously assessed through six complementary stability analyses: manifold analysis, Fisher Information Matrix, Accuracy-Stability Index, data reduction, noise sensitivity, nested cross-validation, and multiple runs with different random seeds. Results reveal a fundamental trade-off between predictive performance and robustness. The deep DeVries18 model achieved the highest accuracy (0.766) and AUC (0.850) with excellent data efficiency—retaining 96.8% of its performance with only 10% of training data—and demonstrated superior generalization in nested cross-validation (AUC=0.780). However, it proved highly fragile against input noise, losing 28.8% of its AUC under SNR=2 perturbation. Conversely, the shallow geometric-feature-based ANN Geo model maintained perfect accuracy stability across all noise levels and exhibited exceptional noise robustness, though with moderate discriminative power (AUC=0.779). The simplified two-layer DNN offered a balanced compromise with 120 parameters, achieving competitive performance (AUC=0.841) and the lowest Fisher condition number (3.65×10³), while the DNN Geo model was rejected due to near-random AUC performance. These findings demonstrate that while architectural depth enhances discriminative capability and data efficiency, it introduces fragility to input perturbations, whereas shallower geometric architectures prioritize robustness at the cost of sensitivity. The results underscore the importance of application-driven model selection in seismic hazard assessment, where data quality and operational constraints should guide architectural choices rather than performance metrics alone.