Utilizing soil–water characteristic curve parameters in custom artificial neural network models to predict the unsaturated hydraulic conductivity

Abstract

Accurately predicting unsaturated hydraulic conductivity (k<inf>u</inf>) is essential for applications in agriculture, environmental management, and geotechnical engineering. This study introduces a predictive model utilizing a custom artificial neural network (ANN) based on saturated hydraulic conductivity and critical unsaturated soil parameters from Alibrahim and Uygar’s model. A unique feature of this ANN is the implementation of a filter within its architecture, ensuring predictions remain within logical bounds by prohibiting negative outputs, which are not physically meaningful for this problem. This embedded domain-specific knowledge makes the ANN robust for this application, as it restricts predictions to feasible values within the context of hydraulic conductivity. The ANN model was trained and tested on a diverse soil texture dataset, achieving high predictive accuracy, as indicated by R² values for both training and testing sets. Analysis of mean absolute error (MAE) and root mean square error (RMSE) values revealed variability in predictive performance across soil textures, with sandy soils showing higher errors due to their unique hydraulic properties and testing method discrepancies. The findings underscore the ANN model’s effectiveness in predicting the unsaturated hydraulic conductivity while highlighting the need for texture-specific adjustments and calibration to address testing-method biases. This approach offers a scalable, data-driven solution for hydraulic assessments in soil hydrology and related fields. © The Author(s) 2025.