A new algorithmic approach for predicting the particle size distribution of dispersed soil suspensions using an automated optical settling column

Abstract

Analyzing a large number of samples for their dispersive properties and particle size distribution (PSD) is often cumbersome and time-consuming, particularly when using traditional hydrometer or pipette-based methods. In this study, we analyzed the efficacy and rapidity of an automated, optical settling column (OSC) to measure the PSD of soils. The OSC provided backscattering and transmission intensity profiles at preprogrammed time intervals that illustrated sedimentation/flocculation processes of suspended particles, making the approach desirable for soil dispersion studies. Settling experiments were performed using 177 soil samples collected from 59 sampling locations across the United States. All soil samples were pretreated with sodium hexametaphosphate (HMP) to achieve maximum dispersion and analyzed with the OSC. The effects of settling run time and sampling frequency were also investigated. The results were compared with separate soil PSD data obtained using the common particle sizing analysis, the pipette method. Using feature engineering and dimensionality reduction, we created different novel methods of modeling the PSD data. Resampling was employed to compare accuracy of these novel methods by generating 300 instances of model performance for each method. We further estimated model generalization using two techniques: (i) a holdout (HO) estimation approach to suggest the “best-case” performance and (ii) an averaged model performance to provide a more conservative performance estimate. The best method studied here provided PSD values that deviated from those gathered through the pipette method by between 2.8% − 3.3% for clay, 5.7% − 9.4% for silt, and 6.5% − 9.6% for sand.

Keywords

• Machine learning
• Particle size distribution
• Soil dispersion
• Feature engineering