Abstract: A model of estimating the maximum dry density of gravel soil is established to provide reference for controlling the compaction quality of gravel soil projects and selecting the gravel soil which meets engineering requirements. In the light that particle gradation is the crucial factor that determines the maximum dry density of gravel soil, 92 groups of data of gravel soil are collected and obtained, of which full gradation (d₁₀-d₁₀₀) is used as the input variable of back propagation (BP) neural network. Furthermore, genetic algorithm (GA) is adopted to optimize the initial weights and thresholds of the BP neural network, based on which the estimation model for maximum dry density of gravel soil is constructed. In addition, the GA-BP neural network model is compared with BP neural network model. According to estimation results, the mean relative error of the predicted results of 86 groups of training samples is 0.54%, and the coefficient of determination is 0.983; the mean relative error of the predicted results of 6 groups of test samples is 0.57%, which indicates that the proposed model is of good generalization performance. It is concluded that the maximum dry density of gravel soil could be well predicted by applying GA-BP neural network in consideration of full gradation. GA-BP neural network model is superior than conventional BP neural network model in terms of convergence rate, prediction accuracy and generalization performance.

Key words: gravel soil, maximum dry density, full gradation, GA-BP neural network, genetic algorithm