Optimization method for epoxy concrete formula based on GA-BP neural network

This study addresses the critical challenge of optimizing aggregate structures to enhance the compressive strength of epoxy resin concrete (Resin E-51 content: 14%) under uniaxial loading. Traditional experimental methods for analyzing multi-scale aggregate interactions are computationally intensive and lack predictive capability. To resolve this, we developed a hybrid computational framework integrating numerical simulation with machine learning. First, gray relational analysis quantified the influence of aggregate parameters on mechanical performance, revealing significant correlations (relational coefficients > 0.5) between fractal dimension, gradation, and strength. Subsequently, a genetic algorithm-optimized backpropagation (GA-BP) neural network with 5–12–1 architecture was established, where genetic algorithms optimized initial weights and thresholds to overcome local minima limitations of conventional BP models. The GA-BP model demonstrated superior predictive accuracy, reducing average compressive strength prediction errors by 78.2% (from 7.0154 MPa to 1.5287 MPa) compared to standalone BP networks. Systematic parameter optimization identified the ideal aggregate configuration: coarse aggregate size = 10 mm, medium aggregate = 5 mm, fine aggregate = 2 mm, fractal dimension = 2.55, and total aggregate content = 85%. These findings provide a mechanistic understanding of aggregate-structure-property relationships while offering an efficient machine learning-driven framework for designing high-performance resin-based composites, with direct implications for infrastructure and construction material optimization.