PSI - Issue 70

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

428

owing to its complementary properties of toughness and crack resistance. Excessive percentage does not contribute greatly to the strength while optimum incorporation does and its observed to be 20%. The experimental results obtained are used in predicting the strength of concrete by the following machine learning techniques. 3. Proposed Machine Learning Model The proposed work utilizes the graph neural network (GNN) for predicting the compressive strength, flexural strength and split tensile strength of SCM under the influence of concrete performance (Ravikanth et al., 2024). The major reasons for utilizing GNN in strength prediction are as follows - Possible to identify the complex interactions among mix components, Capture the heterogeneous inputs in case of mixed SCM components (Moradi et al., 2022). Fig. 3. clearly portrays the working methodology of the GNN based strength prediction. The following section summarizes the implementation of GNN in predicting the strength properties of concrete

Fig. 3. Working methodology of the GNN

3.1. Data preparation and graph construction: The Graph is constructed by identifying the nodes (vertices) and edges (connections). SCM materials, concrete properties and mix ratio are placed over the nodes, whereas the dependencies and other interactions are represented via the edges The SCM materials such as steel fiber (SF - 10%, 20%, 30%, 40%), polyester fiber, Silica Fume, Fly ash (FA - 10%, 20%, 30%, 40%), Metakaolin (MK - 10%, 20%, 30%, 40%) and Rice husk ash (RHA - 10%, 20%, 30%, 40%), concrete properties (compressive, split tensile and flexural strength) under environmental factors are created as nodes. Edges include the different interaction, relationship between strength and durability, cement and SCM, water and SCM., SCM and fiber, steel and poly fiber. Edges denote the connection among the specimens with respect to material configuration and strength over time. One hot encoding is applied for the node attributes such as controlled - 0, SF - 1, FA - 2, MK - 3, RHA - 3. Edges are created between the strength over days such as 7D, 14D, 21D. The input vectors for the compressive strength look like SF 10%(0), 7D, 14D, 21D. Similar type of vectors are generated for split tensile strength and flexural strength 3.2. GNN model development: The type of SCM, proportion and age are passed as input variables. Strength predictions act as target variables. Every node contains 5 input variables, 32 hidden channels and 1 output channel. A typical GNN model includes several layers such as - input layer, graph convolution layer (GCL), pooling layer, fully connected layer and output layer. The nodes and edges from the input dataset are passed onto the input layer. A message-passing Graph Neural Network (GNN) framework is added to model the node - edge interconnections and determine the resulting strength. At each layer l, the feature vector ( ) of a node i is updated by combining the node information from neighboring nodes. The final rule that updates the sum of all neighboring nodes is as follows: