PSI - Issue 70

R. Ashwathi et al. / Procedia Structural Integrity 70 (2025) 424–431

430

Fig. 4. Steps involved in GNN model training The graph dataset is passed as input with the defined set of layers as part of forward propagation. Individual predictions are carried out on all the nodes. The predictions are compared with ground truth annotations using Mean Square Error (MSE) loss function. During the backpropagation, the model weights are updated with Adam optimizer. During hyper parameter tuning the parameters such as - learning rate, batch size, drop out and number of layers are updated for improved model performance. As part of model validation, test data is passed to the GNN model and the results are compared with traditional ML techniques like SVM and decision tree, which is then validated using the techniques such as Mean Absolute Error (MAE), Root Mean Square Error (RMSE) and R 2 (coefficient of determination). The model results are compared with actual results. The following table indicate the statistical differences of the various GCN model. The results of several GNN architectures for forecasting concrete compressive strength show remarkable differences in the accuracy of the model. The conventional GNN model with two hidden layers attained a mean squared error (MSE) of 2.72 and an R² score of 0.79, denoting a rational yet nominal prediction capability. In an attempt to decrease the error rate, increasing the depth to three hidden layers did not support improvements; instead, the model suffered a deterioration in performance with an increased MSE of 3.53 and a slightly better R² of 0.83. This implies that increasing the number of layers may introduce overfitting or suffer from optimization challenges such as vanishing gradients when there is an absence of regularization. In contrast, the GCNConv-based model, which follows graph convolution to combine and bypass information across the graph networks, shows considerable performance when compared to other GNN versions.. It achieved the lowest MSE of 1.18 and MAE of 0.82, alongside an RMSE of 1.09 and an outstanding R² score of 0.98. These results underscore the effectiveness of graph convolution in identifying the complex relationships among concrete mixture components. The high predictive accuracy of GCNConv portrays its agility as an efficient framework for dynamic regression tasks in materials science, particularly in capturing the inherent nature of cementitious composites. 4. Conclusion: The research conducted by different cementitious materials with different percentages at different curing periods poses a significant role in establishing sustainability. The optimal percentage of replacement was observed at 30%. The model is developed using the experimental results attained. The proposed GNN predictive framework depicts greater characteristics for predicting the concrete strength properties. The predicted strength results from the laboratory are passed as inputs to the GNN model. On successful execution, the results are validated against different statistical parameters such MSE, RMSE, MAE and R 2 . Results indicated that GCNConv outperforms the other models with the values 1.18 (MSE), 0.82 (MAE), 1.09 (RMSE) and R 2 (0.98). It is also estimated that increasing the number of GNN layers improves the model performance. Thus the proposed predictive framework can help in strengthening supplementary cementitious materials (SCMs) and fiber combinations, towards increased strength, durability, and sustainability in the latest construction. In future, the dataset can be expanded with increased SCMs, leading to real time capturing and facilitating cost-effective and sustainable material innovation in civil engineering.