PSI - Issue 78

Franco Ciminelli et al. / Procedia Structural Integrity 78 (2026) 921–928

926

in 23% of cases, between 20 m and 50 m in 69%, and > 50 m in 8% (parameter 12). Road alternatives are available for 92% of the bridges and unavailable for 8% (parameter 13). The bypassed entity is of low relevance in 61% of cases, medium in 31%, and high in 8% (parameter 14). Only 5% of the bridges experience frequent pedestrian traffic, while the remaining 95% do not (parameter 15). Finally, 83% of the structures serve a strategic function, whereas 17% do not (parameter 16). Regarding the Seismic Attention Class (CoA- S, parameter 17), 36% of the bridges are classified as “Medium High,” 34% as “Medium,” 14% as “High,” 10% as “Medium - Low,” and 6% as “Low.”

Fig. 2. Learning curves for different dataset sizes during pre-training, showing training and validation metrics for CoA-S prediction.

3. Results This section presents the results obtained in predicting the Seismic Attention Class (CoA-S), using both synthetic data and real data from existing bridges. Fig. 3a shows the confusion matrix derived from the synthetic data, while Fig. 3b illustrates the model performance on real data, using the previously saved model. The corresponding classification reports are provided in Table 2 and Table 3. For the synthetic data, the model achieved an overall accuracy of 0.972, with a macro-averaged F1-score of 0.910 and a weighted F1-score of 0.972. The performance is particularly high for the "High", "Medium-High", and "Medium" classes, with F1-scores of 0.993, 0.967, and 0.960, respectively. The "Medium-Low" class is also predicted with good results (F1-score = 0.878). Performance drops for the "Low" class (F1-score = 0.750), likely due to the limited number of samples (support = 28, being "support" the number of actual occurrences), which affects the model generalization ability. When applied to real bridge data, the saved model achieved an overall accuracy of 0.790 , with a macro-averaged F1-score of 0.762 and a weighted F1-score of 0.783. The model performs well for the "Medium" class (F1 score=0.868) and satisfactorily for the "Medium-High" (0.776) and "Low" (0.857) classes. However, lower performance is observed for the "Medium-Low" (F1-score = 0.667) and "High" (F1-score = 0.643) classes, suggesting the need for further improvements in predicting less represented categories. In summary, the model demonstrates excellent predictive capability on synthetic data, especially for the central classes. On real data, the generalization ability remains good, though improvements are desirable to ensure more consistent performance across all classes, particularly at the extremes of the scale. 4. Conclusions and future development This study has demonstrated the effectiveness of an AutoML-based framework for predicting the Seismic Class of Attention (CoA-S) of existing bridges using only parameters obtainable from census data or technical documentation. The model, trained on synthetic data consistent with the logic defined by the Italian Guidelines, achieved near-perfect performance on these data, with a weighted average precision of 97.2% and an average F1-score of 91.0% .