ÌìÃÀ´«Ã½

A new Artificial Intelligence (AI)-driven study involving ÌìÃÀ´«Ã½ has uncovered hidden processes that challenge the way scientists have modelled how water is retained in soil for half a century. This discovery and new framework could significantly advance water management, irrigation strategies and flood risk assessments worldwide.

The research, undertaken by institutions including Imperial’s Department of Earth Science and Engineering (ESE) and the Department of Agroecology at Aarhus University, reveals that a key assumption used in models of how soil retains water is incorrect.

This finding has implications extending far beyond just the study of soil itself.

When models get the physics wrong, errors can cascade and affect the reliability of a multitude of forecasts – whether that’s for drought, irrigation or even climate simulations. Dr Ben Moseley Assistant Professor in AI, Department of Earth Science and Engineering, Imperial

“The retention of water in the soil governs everything from crop productivity to how landscapes interact with the atmosphere, so when models get the physics wrong, errors can cascade and affect the reliability of a multitude of forecasts – whether that’s for drought, irrigation or even climate simulations,” explained study author , Assistant Professor in AI in ESE.

By developing a hybrid computer model that uses AI to make predictions grounded in physics rules, rather than convenient (but simplified) guesses, the team have overturned a key paradigm in soil science that may have real-life consequences for modelling major processes and systems on a global scale.

A stubborn soil problem

When water sits in soil, it doesn’t all behave the same way: some of it fills the tiny gaps (pores) between the soil particles while some of it clings to the surface of the soil as thin films. Knowing how much of each type of water is present at any given moment provides vital insight into when plants will wilt, how quickly water evaporates from the soil, or whether rainfall will either soak into the soil or run off.

However, elucidating these water types – pore water or film water – is not so simple. Each type is governed by different physical forces, and these processes are notoriously difficult to measure independently at a microscopic scale.

Traditional models of predicting soil water behaviour have therefore had to make educated guesses, relying on the simplified assumption that as soil absorbs more water, the relationship between pore water and film water follows a straight, predictable line.

While this makes the maths easier, it presents another issue in that different models, even when applied to the same soil, are likely to produce drastically different results about how much or little water is available.

A hybrid approach

To improve the process, the study, published , took an unconventional approach to this long-standing problem: combining the basic, known rules that govern how water moves and sticks to surfaces with AI-driven predictions that fill in the gaps.

The team trained an AI model known as a neural network on measurements from 482 soil samples taken from across Central Europe, covering a wide range of soil types. The AI then learnt, from scratch, the patterns that connect a soil’s basic properties (such as its texture and organic matter content) to how it holds water – modelled by what is known as a soil water retention curve (SWRC).

“Our approach marks a departure from conventional methods by embedding a neural network to figure out unknown or poorly understood aspects of SWRCs,” said lead author Sarem Norouzi, Visiting PhD student in ESE, from Aarhus University.

“At the same time, by checking it against physical principles, we made sure that any SWRCs predicted by the model are firmly grounded in reality, despite the AI learning the patterns entirely on its own.”

Soil science breakthrough

Not only did the model accurately recreate SWRCs across the different soil types, it also uncovered hidden connections that traditional models have missed until now. This led to a key discovery: that the transition between soil holding pore water to film water isn’t linear at all.

What makes this significant is that current simplifying assumptions are frequently used in models that inform real decisions about soil and water applications... If the underlying physics isn't represented accurately, the predictions they generate may also be less reliable. Sarem Norouzi Visiting PhD student from Aarhus University

Instead, it turns out that the shift is distinctly curved and more complex than scientists initially realised – meaning that decades of environmental and agricultural models may have been operating with a structural bias.

“What makes this significant is that current simplifying assumptions are frequently used in models that inform real decisions about soil and water applications – from farming and food security, to weather prediction and natural disaster prevention,” said Norouzi.

"If the underlying physics isn't represented accurately, the predictions they generate may also be less reliable.

“Encouragingly, more recent measurements in the lab showed that the transitions predicted by our model were a closer match to the real behaviour than those predicted by state-of-the-art physics-based models.” 

Blueprint for future frameworks

The team’s next priority is to embed their newly discovered soil physics into the broader hydrological and climate models that currently rely on the old assumptions, presenting a considerable task of updating the systems used by scientists and policymakers worldwide.

The researchers are also extending their framework to model hydraulic conductivity (how easily water moves through soil), which is closely linked to water retention and equally critical for accurate predictions. They plan to test their approach across additional soil databases and field conditions to ensure it generalises beyond the European samples used in this study.

“A final, striking outcome of our work is the demonstration of a new way of doing science,” said Dr Moseley. “We’re moving away from the old paradigm of prescribing how nature should work and towards one where we let nature tell us how it does work, to reveal how complex environmental systems actually behave and to rebuild our foundations on firmer ground.”