Predicting concrete compressive strength using optimized deep learning and large language models

Concrete compressive strength prediction presents a fundamental challenge in sustainable construction material design due to the intricate nonlinear interactions among mixture components, admixtures, and curing conditions. This study introduces a hybrid framework that integrates the cognitively inspired iHow Optimization Algorithm (iHowOA) with Spatio-Temporal Graph Convolutional Networks (STGCN) to enhance predictive accuracy for concrete compressive strength estimation. A large language model (LLM)-driven preprocessing pipeline is employed to improve data quality through semantic validation, feature harmonization, and intelligent handling of inconsistencies, leading to robust input representations. The iHowOA optimizes the STGCN architecture by leveraging hierarchical knowledge acquisition, balanced exploration-exploitation, and adaptive decision-making mechanisms. The graph-based model captures spatial dependencies among compositional variables and temporal strength evolution during curing. Extensive benchmarking against ten established metaheuristic optimizers demonstrates that the proposed iHowOA-STGCN framework achieves superior predictive performance on the evaluated public dataset, yielding lower prediction errors and higher correlation coefficients compared to baseline models. Exploratory data analysis further highlights key cement-strength relationships, age-dependent strength gain patterns, and physicochemical interactions relevant to feature engineering. While the experimental results indicate that the proposed framework is a promising data-driven decision-support approach for concrete strength prediction, further validation on diverse datasets and real-world scenarios is required to assess its generalizability and practical applicability.