Aviation safety depends on accurate identification and effective prioritization of operational hazards. The ICAO risk matrix, despite its simplicity and global acceptance, has limitations in identifying complex and borderline scenarios. This article compares three risk matrix models the standard ICAO matrix, the dual-filter matrix, and the modified resilient matrix in terms of learnability by a neural network and validation with real data. Unlike previous studies, 47,723 real event reports from the NASA database and five input features were used with a multilayer perceptron network. Model performance was evaluated using MSE, RMSE, MAE, and R² criteria. The modified resilient model showed the best performance with R²=0.9612 and MAE=1.108; the dual-filter model and the standard ICAO matrix ranked next. The Friedman test confirmed the significant superiority of the neural network (p < 0.001). Adding intermediate levels to the ICAO matrix, without changing the logic of multiplying probability and severity, improves learnability. The dual-filter model, despite its good performance, is not recommended for online applications due to computational complexity. Validation with real data has shown that there is a significant gap between model performance on simulated and real data, and the use of real data in future research is essential.