Artificial intelligence-based prediction and optimization of textile waste recovery in hydrated plaster: A machine learning and meta heuristic approach

This study investigates the creation of eco-efficient gypsum plaster composites that integrate untreated textile waste (TW) with a hybrid experimental-computational methodology that combines machine learning (ML) prediction with multi-objective metaheuristic optimization. Prismatic specimens of 40 × 40 × 160 mm were fabricated with TW concentrations varying from 0% to 1% and water-to-plaster (W/P) ratios between 0.55 and 0.70. The composites were evaluated for rheological parameters (initial and final setting times, spreadability), durability (capillary absorption), mechanical performance (compressive strength [CS] and flexural strength [FS]), and thermal conductivity (TC). The results indicated that TW markedly affected plaster performance: a 0.75% TW addition produced the maximum compressive strength (11.67 MPa) and flexural strength (4.17 MPa), while thermal conductivity reduced from 0.20 to 0.15 W/m·K, hence improving thermal insulation. Nonetheless, workability was diminished—spreadability decreased from 210 mm to 130 mm, and initial setting time reduced from 7 to 3 min—underscoring a trade-off. A deep neural network enhanced by the Improved Grey Wolf Optimizer (DNN–IGWO) attained superior prediction accuracy (R² > 0.95), proficiently simulating nonlinear relationships between TW and W/P ratios. A genetic algorithm (GA) produced a Pareto front of 71 non-dominated solutions, optimizing strength, thermal performance, and workability. Optimal formulations were achieved at W/P ratios of 0.55–0.65 and TW levels of 0.25–1 wt%, enabling the development of high-performance, sustainable gypsum composites derived from industrial textile by-products. The findings support the incorporation of recycled textiles in construction and illustrate how data-driven optimization can inform the advancement of sustainable gypsum-based materials utilizing industrial by-products.