PSI - Issue 70

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

425

1. Introduction The rapid urbanization has caused the demand towards high strength concrete due to erection of modernized structure and skyscraper buildings (Kang et al., 2024). This in turn creates an urge towards alternative building materials and sustainable materials as the aggregates are natural resources (Mohanraj and Vidhya, 2024). On the other hand, cement used as a binding material contributes towards increased emission of carbon footprint. It is equally essential to find an alternative for both aggregates and cement to establish the triple bottom line of sustainability. One such promising solution is utilization of Silica Fume, Fly ash, Metakaolin and Rice Husk Ash, the supplementary cementitious materials in concrete for cement where it is partially replaced at 10%, 20%, 30% and 40% (Bakera Mark, 2019). These materials help in reducing greenhouse gas emission and also exhibit pozzolanic behavior, an essential property in enhancing the performance of concrete (Loganathan et al., 2022). It contributes towards improved strength by reduction in porosity, and strengthening the Interfacial Transition Zone (Golewski, 2023). The addition of SCMs in concrete mixes is a regular procedure as it results in decreased cost of production, improved performance and reduction in environmental degradation. The conventional procedure of predicting the performance of BCC incurs immense laboratory testing that can be expensive and time-consuming. In order to overcome the drawbacks of conventional methods, researchers created predictive models with the help of machine learning algorithms (Padmapoorani et al., 2023). ML algorithms are considered to be time-saving and efficient techniques in estimating concrete strength prediction under the influence of SCM. Decision trees and Adaboost regressor model were used in predicting the compressive strength and durability of BCC (Güçlüer Abdurrahman; Göymen, Samet; Gunaydin, O., 2021). The dataset consisted of a combination of 1900 data points for compressive strength, chloride penetration and carbonation. Results indicated that both the models achieved a correlation coefficient exceeding 0.9 in both the predictions. Statistical error analysis revealed a value of less than 0.5 chloride resistivity and carbonation (Moradi et al., 2022). ANNs are used in predictive modeling due to its principal properties such as coherent handling of non-linear structures, adaptability to diverse constitution of materials, increased forecast accuracy, cost effective and improved generalization (Prasad Hamid; Reddy, B. V. Venkatarama, 2009). ANN based research was carried out in concrete strength prediction under the influence of binary SCM with a collection of 19 dissimilar features. Results indicated an overall error rate less than 10% ensuring higher accuracy. The overall error rate was calculated using different statistical error analysis methods such as Mean Absolute Percentage Error, Root Mean Square, Nash- Sutcliffe Efficiency and R (correlation coefficient). In a similar manner, ensemble methods such as random forest, Adaboost, light-weight gradient machine (LGM), gradient boosting machines (GBM) offer great performance when compared to ML models, due to their ability of combining several weak learners aka base models (Chopra Rajendra Kumar; Kumar, Maneek; Chopra, Tanuj, 2018). The ensemble models were trained on two datasets and experimental results indicate that XG boost outperforms the several ML and other ensemble models with an accuracy around 90% and decreased statistical errors. Bayesian optimization was applied to certain ensemble models for better hyperparameter optimization resulting in decreased parameter search time and better accuracy. One such model was an explainable boosting machine (EBM), gradient boosted regression tree (GBRT) that utilizes tree based integration algorithms for improving prediction accuracy (Ahmad Furqan; Ostrowski, Krzysztof; Śliwa - Wieczorek, Klaudia; Czarnecki, Sławomir, 2021) . Though ensemble models offer a noteworthy performance, there are certain limitations which remain unaddressed. The ensemble model works well for static modalities, lacks in generalization and captures very intrinsic feature representations. The proposed problem statement considers the composition of several SCM’s and there is a need for a learning model that can learn dynamic features of the novel SCM proportions. Graph neural network (GNN) works well for dynamic feature learning and adapts to different types of data structures based on the given modality. Nomenclature SF Silica Fume FA Fly Ash MK Metakaolin RHA Rice Husk Ash