A practical approach for soil unit weight estimation using artificial neural networks

Abstract

The soil unit weight is a relevant geotechnical and geological parameter in interpreting field tests, estimating parameters, and developing designs. The most precise way to estimate the soil unit weight is through controlled laboratory tests, which require samples with undisturbed structures. However, such tests produce results specific to the extracted sample point and depth, making the abstraction difficult in large areas. Efforts in the literature rely on in situ penetration tests and mathematical equations to simplify estimating the parameter value for general scenarios. On the other hand, each sort of soil can have a characteristic behavior, reducing the precision of estimations and hindering the adoption of statistical-based models universally. We propose a practical approach supported by machine learning to predict the soil unit weight, which produces estimations with an R2 of 0.82. The methodology consists of three phases: (i) we perform a clustering analysis to understand if different soils have similarities that permit the generation of multiple models; (ii) we statistically examine soils parameters to identify relevant parameters and produce a regression model, and; (iii) we feed an ANN with in situ parameters to estimate the soil unit weight. We make publicly available a web application to estimate the soil unit weight.