Deep neural network modeling of the properties of sustainable high-performance concrete from industrial waste materials

Abstract

• Adding recycled metal fibers improves the physicomechanical properties of high-performance fiber-reinforced concrete (HPFRCs). • Carrying out sensitivity assessment analysis for each constituent of the reinforced mixture. • HPFRC curing time considerably influences its performance. • The newly developed deep neural network models are highly accurate for predicting HPFRC flexural strength, compressive strength, tensile strength, and water absorption. • HPFRC algorithm optimization and engineering requirements are clearly described. The manufacture of concrete significantly impacts the environment due to the substantial consumption of non-renewable resources and CO₂ emissions generated during cement manufacturing. This work develops a sustainable high-performance fiber-reinforced concrete (HPFRC) by integrating recycled stainless-steel fibers (SFC) sourced from industrial cables, thereby advancing circular economy principles. An extensive experimental study was performed to evaluate the effects of variations in water-to-binder ratio ( W/B ), fiber content (SFC), fiber aspect ratio ( L/d ), and curing time ( T ) on four principal properties: compressive strength (CS), flexural strength (FS), splitting tensile strength (STS), and water absorption (WA). The optimal mechanical performance was attained with a mix design of W/B = 0.29, L/d = 63, SFC = 29 kg/m³, and a curing time of 90 days, resulting in CS = 115.36 MPa, FS = 11.85 MPa, and STS = 9.46 MPa. The least water absorption (0.42 %) was recorded with a water-to-binder ratio of 0.27, a length-to-diameter ratio of 63, and a specific gravity of 24 kg/m³ at 90 days, signifying exceptional durability. Analysis of variance (ANOVA) indicated that curing time exerted the most substantial influence on all mechanical parameters, accounting for up to 69.6 % of the observed variance. Six deep neural network (DNN) architectures were constructed to represent and forecast these qualities, with each architecture optimized using a distinct algorithm: Genetic Algorithm (GA), Dragonfly Algorithm (DA), Improved Grey Wolf Optimizer (IGWO), Levenberg–Marquardt (LM), BFGS, and Conjugate Gradient (CGP). The IGWO-DNN model demonstrated superior predictive ability, attaining R² values over 0.98 for all outputs, accompanied by negligible prediction errors (MAPE of 0.53 % for CS and 1.33 % for WA; RMSE of 0.88 MPa for CS and 0.20 % for WA). This combined experimental-AI framework illustrates an effective method for designing eco-efficient concrete with enhanced mechanical and durability properties, utilizing industrial waste and sophisticated optimization algorithms.