AI models a surprising shift in how drying soil holds water

A machine-learning model suggests that, as soil dries, water held in surface films can initially increase even while total water falls. The study, published earlier this year in the Journal of Hydrology, proposes a physical explanation for that counterintuitive transition.[1]

The explanation begins with the soil's architecture. Water occupies spaces between particles, including corners where curved interfaces help retain it. Water also adheres to mineral surfaces in very thin films. Capillary effects dominate wetter conditions; the films become especially important at the dry end. Treating soil as a bucket with a single amount left in it misses how differently that water is held.[2]

Markus Tuller and Dani Or explored this distinction in experiments published in 2005. They found that accounting for the mineral surface area brought the dry-end behaviour of different soils into much closer alignment. A little water can coat a great deal of surface. What looks like the last, uninteresting residue is therefore carrying information about the material that holds it.[2]

A curve made from real soil

Soil scientists describe the relationship with a water-retention curve: the amount remaining as the strength of suction changes. Such curves help connect basic soil properties to drainage, evaporation and water exchange with the atmosphere. Their useful range stretches from saturation to extreme dryness, while the observations available for learning are often clustered toward wetter conditions.[4]

The 2026 investigation drew on a substantial laboratory resource published in 2023: 572 sets of soil properties, mostly from central Europe. Its samples were collected in rigid cylinders that preserve soil structure. Different techniques covered different moisture ranges, combining evaporation measurements with a dewpoint method for drier material. A complete curve therefore rests on several kinds of measurement, rather than a sensor directly watching every microscopic store of water.[3]

This is an unusually useful setting for scientific AI. The data describe what the whole sample does. The question concerns the processes inside it. A model can be evaluated against the measured total, while the interpretation of its internal parts requires a further argument. Keeping those two jobs visible is essential when a prediction is also being offered as an explanation.[3]

Keep the physics that is known

Earlier work by Sarem Norouzi and colleagues showed why a neural network benefits from physical structure. In their 2025 study, a conventional network could produce curves that turned the wrong way or behaved unrealistically outside densely measured regions. A constrained version preserved the decrease in water content as suction increased and imposed suitable behaviour at the extremes. These restrictions helped the model draw a sensible curve where examples were sparse.[4]

Using 482 eligible samples from that collection, the 2026 hybrid combines an analytical description of dry soil with neural networks learning wetter behaviour and the transition between regimes. Its inferred transition is nonlinear. As pores drain, additional surfaces become exposed, so water assigned to films can initially increase before those films thin. This can happen while total water steadily falls.[1]

For held-out samples, the reported root-mean-square error was 0.049 cubic metres of water per cubic metre of soil. Testing used 143 samples and 2,478 measurement points. All observations from the same soil profile stayed together when training and test data were separated.[1]

The research group's description emphasises learning uncertain physics by adding machine-learning components to an existing model. That is the useful design idea here: preserve established relationships, expose the part that remains uncertain and ask whether the observations can constrain it.[5]

The predicted division between capillary and film water still needs independent validation. Agreement with total water does not establish that both hidden contributions are correct. The payoff is a physically constrained hypothesis about where the water goes as soil dries, with a specific transition for further experiments to investigate.[1]

Report an error or suggest a correction