Modeling standard lines for soil compaction testing using artificial neural networks and geometric algorithms

Abstract

Soil compaction testing is crucial for ensuring the stability and durability of infrastructure projects. Traditional methods for generating standard lines, such as averaging and polynomial fitting, often fail to capture the nonlinear relationships and variability in compaction data, leading to inaccuracies in soil property assessments. This study introduces a novel framework that leverages Artificial Neural Networks (ANNs) to dynamically model standard lines for compaction curves, addressing limitations of traditional approaches. Four activation functions—ReLU, Sigmoid, Tanh, and Swish—were evaluated, with Swish emerging as most effective for capturing complex relationships between Dry Density (DD) and Moisture Content (MC). A tolerance-based evaluation framework, incorporating tolerance levels of 2%, 5%, and 10%, was applied to analyze coverage areas. The 5% tolerance level was identified as most balanced, minimizing errors while providing reliable representations of compaction data. The study also introduced the Ray-Casting Algorithm for precise calculation of coverage areas, enabling a new performance indicator based on density of data points within the region. Results demonstrate that the ANN framework, particularly with Swish activation, outperforms traditional statistical methods in accuracy and adaptability. ReLU delivered the best performance, with the lowest prediction and percentage errors (0.1910–0.2005 and 10.66%–11.74%), while effectively explaining over 55% of the data’s variability. Sigmoid showed the weakest results, with high errors and near-zero variance explanation. Tanh performed moderately, balancing accuracy and generalization with reasonable error levels and 44%–54% variance capture. Swish was consistently reliable, with stable errors and over 50% of the variance explained. This research advances compaction testing by addressing variability, operator-induced errors, and nonlinear data patterns, establishing a reliable methodology for generating standard lines. Future work could explore diverse soil types, integrate environmental factors, develop hybrid machine learning models, and improve performance indicators.