Impact Analysis and Prediction Research of Soil-water Characteristic Curves Based on Data Mining

Soil-water characteristic curves of unsaturated soils are affected by many factors.It is critical to reveal the active characteristic of each factor and ascertain the impact mechanism of the dominant factors for unsaturated soils in various conditions.To remedy this issue, mathematical statistical analysis and machine learning methods are employed to analysis the sensitivity of soil-water characteristic curve to each factor by investigating the variation of three characteristic values: Air-entry value, dehumidification rate and residual water content, in which the basic data is collected from the existing literatures.Then, Genetic Neural Networks(GA-BP) is introduced to predict those characteristic values of soil-water characteristic curve in consideration of the impact of multiple factors.The results show that:(1) Spearman correlation coefficient results verified that at least eight factors affect the three characteristic values in different degrees.(2) By partial correlation analysis and machine learning, the plasticity index and the dry density collectively dominate the variation of the three characteristic values, and the third factor is the grain composition.In the grain composition, the effect of the effective particle size d₁₀ is most significant.(3) Sensitivities of different characteristic values to a factor are different.The plasticity index influences all three characteristic values of soil-water characteristic curve to a similar degree.The effect of dry density on air-entry value was slightly greater than its effect on residual water content.The effect of grain composition focuses on residual water content, while it is slight on the dehumidification rate.The effect of drying-wetting cycles mainly reflects in the air-entry value and the dehumidification rate.(4) Distributions of the three characteristic values under the effects of dry density and plasticity index are achieved, and also the reference ranges of three characteristic values are suggested for different soils.(5) The GA-BP neural network is valid in the prediction of the three characteristic values for soil-water characteristic curve under the combined action of multiple factors, and the evolution characteristics of soil-water holding performance can be reasonably reflected.