Machine Learning Predicts How Water Moves Through Soil in Milliseconds

Measuring how water flows through unsaturated soil has long been one of the most expensive and slowest tasks in geotechnical and hydrological engineering. Laboratory techniques such as pressure plate apparatus, Tempe cells, and steady-state flow experiments demand specialized equipment, skilled operators, and often months of work for a single sample. Yet the hydraulic conductivity function, which describes how easily soil transmits water as it dries, underpins nearly everything that happens in the vadose zone: infiltration of rainfall, evapotranspiration from crops, the stability of slopes after storms, and the migration of contaminants toward groundwater. A new study published in Discover Artificial Intelligence by Ishwor Thapa of Tribhuvan University and Amrit Poudel offers a machine learning framework, complete with an interactive graphical user interface, that estimates these hydraulic properties in milliseconds from soil data that any field technician can collect.

The core of the framework is a set of ensemble machine learning models trained to predict the five parameters of the widely used van Genuchten formulation: residual water content, saturated water content, the air entry parameter alpha, the pore size distribution parameter n, and the saturated hydraulic conductivity. Once these five numbers are known, the Mualem-van Genuchten equations generate a complete soil water retention curve and a full unsaturated hydraulic conductivity curve, showing how conductivity changes across the entire range of pressure heads relevant to field conditions. The inputs are deliberately simple: sand, silt, and clay percentages, bulk density, organic carbon content, particle size indices such as d10, d50, and d60, the uniformity coefficient, and porosity calculated from bulk density. These are properties that can be obtained from routine soil characterization rather than months of specialized hydraulic testing.

The authors trained two algorithms known for their performance on tabular data: Random Forest, configured with 150 decision trees limited to a maximum depth of 15, and XGBoost, a gradient boosting method using 150 trees with a maximum depth of 5, a learning rate of 0.1, and both L1 and L2 regularization. All input features were standardized before training, and the saturated conductivity target was log-transformed to reduce its strong right skew. Hyperparameters were tuned with randomized search and five-fold cross-validation on the training subset, and final performance was assessed on an independent holdout test set comprising 25 percent of the data, or 375 samples.

The results reveal both the promise and the honest limits of the approach. On the holdout set, the models achieved a coefficient of determination above 0.99 for the air entry parameter alpha, above 0.98 for the pore size distribution parameter n, and above 0.93 for saturated water content. Residual water content was predicted with good accuracy at roughly 0.76. Saturated hydraulic conductivity, however, proved far harder, with an R-squared of only about 0.15. This weakness is not unique to the new framework: benchmark comparisons against multiple linear regression and the established ROSETTA3 pedotransfer model showed that every method struggled with saturated conductivity, achieving R-squared values between roughly 0.12 and 0.18. The authors conclude that the limitation is largely inherent to the input feature set rather than the choice of algorithm, because conductivity varies enormously with soil structure features such as cracks and macropores that particle size and density data simply cannot capture.

An important caveat runs through the study, and the authors state it explicitly: the training data are synthetic. Rather than measuring real soils, the team generated 1500 virtual soil samples whose feature distributions were calibrated to match the summary statistics of UNSODA, a global database of 790 measured soils. Sand, silt, and clay fractions were drawn from a Dirichlet distribution to enforce the physical constraint that they sum to 100 percent, bulk density came from a truncated normal distribution centered at 1.45 grams per cubic centimeter, and organic carbon followed a log-normal distribution. Target parameters were then computed from published empirical relationships, including the classic Cosby equations and ROSETTA3-calibrated regressions, with realistic measurement noise added to mimic laboratory uncertainty. The reported accuracies therefore reflect the models’ ability to reproduce synthetic reference targets, not independent experimental observations, and the synthetic data cannot capture natural structural heterogeneity such as cracks, macropores, or hysteresis present in real soils.

Despite that limitation, the framework’s key advantage over classical pedotransfer functions is its ability to predict all five van Genuchten parameters simultaneously. Traditional empirical formulas, such as the Cosby relationship linking conductivity to sand and clay fractions or the Saxton and Rawls equations, predict saturated conductivity alone and cannot generate complete conductivity curves without additional assumptions. They also exhibit systematic biases. The machine learning approach, by contrast, produces consistent predictions across the full pressure head range, and comparison of predicted curves against synthetic references for representative samples showed excellent agreement for water retention and good agreement for conductivity, particularly in the mid-range pressure heads most relevant to field conditions. The authors caution that because saturated conductivity prediction remains weak, the absolute magnitude of predicted conductivity curves should be treated as a preliminary estimate rather than a reliable measurement.

Feature importance analysis added physical credibility to the models. For saturated water content, porosity dominated with 51.5 percent of predictive importance, followed by bulk density at 43.0 percent, while sand and clay content each contributed only around 1.3 to 1.4 percent. This ranking aligns precisely with physical understanding, since porosity and bulk density directly control how much water a saturated soil can hold, while texture plays a secondary role. Accuracy also varied by texture class: predictions were best for sandy soils, where pore geometry is more uniform, and worst for clay-rich soils, where structure, cracks, and macropores exert a dominant influence that basic particle size and density inputs cannot represent.

The practical deliverable is an interactive graphical user interface released in two versions: a Streamlit-based Python application with the full machine learning backend, and a standalone HTML and JavaScript version that runs in any browser. Users enter ten basic soil properties, the interface validates that texture fractions sum to 100 percent, automatically calculates derived properties such as porosity and the uniformity coefficient, and then displays predicted parameter cards alongside interactive soil water retention and hydraulic conductivity curves. The visualization includes markers for field capacity and wilting point, highlights plant available water, supports batch processing of hundreds of samples via CSV upload, and exports results to CSV or Excel. The authors argue that this accessibility matters: researchers and practitioners without specialized training in soil physics can obtain first estimates that previously required months of laboratory work, at essentially no computational cost.

The study is candid about what must happen next. The highest priority, according to the authors, is independent experimental validation against measured hydraulic parameters from UNSODA or another experimental database, replacing synthetic targets with real laboratory data from the soil types of interest. They also recommend incorporating additional features such as Atterberg limits and mineralogy, implementing uncertainty quantification by propagating prediction intervals through the van Genuchten-Mualem equations, and exploring physics-informed neural networks to enforce physical constraints directly during training. For geotechnical applications, where point predictions may be insufficient given natural variability, quantile-based confidence bands derived from the distribution of individual Random Forest tree predictions could produce confidence envelopes for entire curves. The model’s applicability is explicitly constrained to the feature ranges of the training data, and extrapolation beyond those bounds is not recommended.

For a field where a single hydraulic measurement can consume months and thousands of dollars, the ability to generate complete conductivity and retention curves in milliseconds from routine soil data is a genuinely useful capability, even with its acknowledged limits. The framework positions machine learning not as a replacement for careful laboratory measurement but as a screening tool: valuable first estimates for agriculture, environmental engineering, and preliminary geotechnical design, with critical or high-risk projects still requiring selective laboratory verification. If future versions trained on measured soils and equipped with uncertainty quantification hold up under experimental scrutiny, the approach could become a standard first step in characterizing the hidden plumbing of the ground beneath our feet.

Subject of Research: Machine learning estimation of unsaturated soil hydraulic conductivity functions from basic soil properties

Article Title: Machine learning estimation of unsaturated soil hydraulic conductivity functions using literature-derived pedotransfer functions and an interactive graphical user interface

Article References: Thapa, I., & Poudel, A. (2026). Machine learning estimation of unsaturated soil hydraulic conductivity functions using literature-derived pedotransfer functions and an interactive graphical user interface. Discover Artificial Intelligence, 6(1), Article 1437. https://doi.org/10.1007/s44163-026-02374-x

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02374-x

Keywords: unsaturated soil, hydraulic conductivity, pedotransfer functions, machine learning, Random Forest, XGBoost, van Genuchten model, soil water retention curve, graphical user interface, soil physics, vadose zone, UNSODA