A quantitative evaluation method regarding the natural void ratio of undisturbed loess

Abstract

Objective The natural void ratio is the most frequently used and important characterisation parameter of the initial structure at the macroscopic level. Therefore, the analysis and prediction of the distribution pattern of the natural void ratio of loess is important for understanding undisturbed loess disaster mechanics behaviour and for disaster early warning from the geotechnical point of view. Methods In this study, particle analysis tests, XRD tests, natural void ratio tests and 1D consolidation tests were carried out on in situ soil samples from different layers of a typical loess site to analyse the correlation between the natural void ratio and particle fraction and stress history. The results show that the natural void ratio can be affected by the stress history and particle size distribution. The higher the overburden pressure is, the more uniform the grading is and the smaller the natural pore ratio is. The water content may be one of the reasons for the variation in the natural void ratio. Results On this basis, the burial depth of the layer, the inhomogeneous coefficient and curvature coefficient of particle gradation, and the natural water content are selected as the influencing factors, and the natural void ratio is evaluated quantitatively based on the machine learning algorithm. The SSA and PSO algorithm were introduced to optimise the weights and thresholds of the BP neural network, and natural void ratio predicted models based on the BP, SSA-BP and PSO-BP neural networks were established. The trained BP, SSA-BP and PSO-BP neural network models were then used to predict 16 sets of validation and test data, and the predicted results were compared with the measured natural void ratios. Conclusion The results show that the PSO-BP-based neural network model predicts significantly better than the SSA-BP and BP neural network models, and can effectively predict the natural void ratio.

Keywords

• undisturbed loess
• natural void ratio
• particle gradation
• stress history
• bp neural network