PREDICTING CONCRETE COMPRESSIVE STRENGTH USING LONG SHORT-TERM MEMORY DEEP LEARNING: A DATA-DRIVEN APPROACH FOR ROBUST AND GENERALIZABLE STRENGTH ESTIMATION

Abstract

Accurate prediction of concrete compressive strength is essential for ensuring structural safety, optimizing mix design, and reducing experimental cost and time. However, the inherent nonlinearity and heterogeneity of concrete materials, particularly when supplementary cementitious materials are involved, limit the effectiveness of traditional empirical and conventional machine learning models. This study proposes a Long Short-Term Memory (LSTM)–based deep learning framework to predict concrete compressive strength using a large and diverse dataset comprising 1,133 samples with varying mix proportions and curing ages. The model incorporates cement, blast-furnace slag, fly ash, water, super-plasticizer, coarse aggregate, fine aggregate, and age of testing as input features. Model performance was rigorously evaluated using training, validation, and test datasets and assessed through multiple statistical and error-based metrics, including R², Nash–Sutcliffe Efficiency, RMSE, MAE, MAPE, and CVRMSE. The results indicate strong predictive capability, with R² and NSE values consistently exceeding 0.80 across all data partitions, demonstrating effective learning and generalization. Error analysis revealed that prediction errors are largely centred near zero with unimodal distributions, while only minor underestimation was observed for high-strength concrete. Compared with findings reported in recent literature, the proposed LSTM model shows competitive performance and enhanced robustness when applied to a broad strength range and heterogeneous mix designs. The outcomes confirm the suitability of LSTM-based models as reliable tools for preliminary concrete mix design and performance assessment, offering a promising alternative to labor-intensive experimental testing and conventional predictive approaches.