Prediction of compressive strength and characteristics analysis of semi-flexible pavement desert sand grouting material based upon hybrid-BP neural network

Abstract

Evaluating the mechanical properties of desert sand grouting Material (DSGM) utilized in semi-flexible pavement within practical engineering applications is crucial for ensuring its safe utilization. To precisely obtain DSGM exhibiting exceptional mechanical properties, the Backpropagation Neural Network (BPNN) model was optimized through the utilization of Particle Swarm Optimization (PSO), Sparrow Search Algorithm (SSA), and Genetic Algorithm (GA). Relationships between water-cement (w/c) ratio, desert sand (DS) content, fly ash (FA) content, bentonite (BT) content, and superplasticizer (SP) dosage were established in relation to compressive strength. Experimental flexural and compressive strengths served as evaluative indices for the mechanical properties of DSGM. Correlation matrix analysis and Principal Component Analysis (PCA) were conducted to ascertain the relationships between various raw materials and the mechanical properties of DSGM, while comparative analyses were also performed on these mechanical property evaluative indices.The results indicated a positive correlation between DS content and SP dosage with compressive strength, whereas a negative correlation was observed between w/c ratio, FA content, and BT content with compressive strength. DS effectively dispersed the DSGM cementitious material slurry, leading to a more uniform distribution of hydration products and a stronger bond in the transition zone of the aggregate interface. Consequently, the DSGM matrix structure became more dense, resulting in higher compressive strength. Through PCA, the importance of different variables and their overall scores were analyzed, revealing Group NO12 as the optimal mix ratio. The GA algorithm significantly enhanced the predictive accuracy of the BPNN model. The predictive performance evaluative indices for compressive strength using GA - BPNN were R² = 0.93, MAE = 2.39, MAPE = 0.06, and RMSE = 3.18. Therefore, GA - BPNN demonstrated the highest predictive accuracy for the compressive strength of DSGM, providing novel insights for the mix design of DSGM.