Proposing an optimised data-efficient artificial neural network model for estimating the compressive strength of concrete containing SCMs

The development of innovative, precise, and cost-efficient methods for predicting the mechanical properties of construction materials, particularly concrete, has become an important topic in the field of civil engineering. This research presents a novel method for forecasting the compressive strength of concrete that incorporates supplementary cementitious materials (SCMs) such as Sugarcane Bagasse Ash (SCBA) and copper tailings while optimising data efficiency. Achieving accurate predictions with limited data is challenging, including for eco-friendly materials designed to minimise the environmental impact of concrete manufacturing. Two models were evaluated: the Radial Basis Function Neural Network (RBFNN) and the proposed Adaptive Fine-Tuned Neural Network (AFTNN). The AFTNN employs a two-step process: pretraining on conventional concrete data without SCMs and fine-tuning with limited SCMs data. This method leverages pre-existing information to improve prediction accuracy, even with restricted datasets. The models were run using two main datasets: one containing SCBA and the other comprising Copper Tailing (CPT) data. In the first group, 779 data points were introduced to the model, while in the second group, 611 data points were used. For all models and datasets, the data were randomly divided into three main groups: 70 % for training, 15 % for testing, and 15 % for validating. The results indicate that while the RBFNN performs well in predicting the compressive strength of concrete mixes with SCMs, for the SCBA data in testing, the Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) values are 3.6 and 4.34, respectively. The AFTNN consistently achieves even better accuracy, with RMSE and MAE values of 3.2 and 4 for the same dataset. While reducing training data generally decreases model accuracy, the two-step AFTNN process mitigates this impact by effectively utilising both normal concrete and SCM datasets. For instance, the AFTNN achieved superior results even when the training data was reduced by 40 %, with R-values of 0.93 and 0.925 for SCBA data in the testing and validating phases, respectively. Moreover, the RBFNN yielded R-values of 0.91 and 0.89 under the same conditions. The results for CPT data confirm this trend as well. These findings highlight the capability of the AFTNN to provide reliable predictions for eco-friendly concrete mixes, emphasising its value in sustainable construction practices. This research demonstrates the potential of machine learning tools to overcome data limitations, clearing the path for more effective and sustainable methods in predicting the performance of advanced concrete materials.