Sarem Norouzi from the Department of Agroecology – Soil Physics and Hydropedology at Aarhus University, digs into a new paper he’s worked on in Water Resources Research. It focuses on a new machine learning approach that enhances the Soil Water Retention Curve’s role in soil health.
The Soil Water Retention Curve and its role in soil health
The Soil Water Retention Curve (SWRC) describes the relationship between the water content of soil and the soil suction (matric potential). It essentially shows how much water soil can hold at various levels of suction, from saturation to near complete dryness. This curve is closely linked to soil health because it controls how much water is available to plants, how water and air move through the soil, and the conditions that impact microbial activity.
Data limitations in SWRC measurement and challenges with conventional models
Despite its importance, measuring the SWRC is time consuming and technically challenging. To measure the entire curve from saturation to near dry conditions, we need to use several devices, each for a certain range of suctions. As a result, complete SWRC datasets are rare, and many existing datasets only cover the wet end (lower suctions).
Another challenge is that conventional modelling methods have limitations that prevent the use of many existing SWRC measurements. These traditional models typically assume a specific parametric form for the retention curve and require complete data for each sample for model training. Consequently, datasets with incomplete or uneven measurements are often excluded, resulting in a significant loss of spatial coverage in SWRC mapping and modelling.
Integrating physics and machine learning to predict SWRC
In our recent paper, published in Water Resources Research (Norouzi et al., 2025), we proposed a novel physics-informed machine learning approach that estimates the entire continuous SWRC using predictors such as soil texture, organic carbon, and bulk density. The key advantage of this method is its ability to overcome the limitations of traditional models, which often require complete and balanced datasets. We developed a physics-informed model that integrates established physical principles directly into the learning process. The core idea is to guide the model not only with data but also with known physical behaviours of the SWRC. By embedding these physical constraints, the model can make robust and physically plausible predictions, even when data is incomplete or unevenly distributed.
This approach offers several advantages over conventional methods: it allows the use of incomplete data sets that would otherwise be discarded, ensures physically consistent predictions, and enables better generalisation across diverse soil types. Moreover, because the model produces a continuous, differentiable SWRC, it can be directly integrated into hydrological and soil-plant-atmosphere models.
Application of physics-informed neural network to predict the soil water retention curve. (a) The approach allows the model to be trained on soil water retention curve on (1, 2) very few points or (3) measurements in the dry range only. (b) The model was applied to the whole of Denmark (inset) with the shown comparison of two predicted soil water retention curves with different clay contents. Figure prepared with contributions from Prof. Peter Lehmann.
Towards a European map of Soil Water Retention Curves
Our work on the SWRC has been successfully applied to map the full soil water retention curve across Denmark. The physics-informed method allowed us to include all available samples in the training process, avoiding the typical loss of spatial coverage caused by model limitations or data gaps. We are now collaborating with partners from the AI4SoilHealth project to scale this approach across Europe. The goal is to develop a new map of the SWRC for Europe, along with related factors, such as plant available water and other key indicators of soil hydraulic function.
References
Norouzi, S., Pesch, C., Arthur, E., Norgaard, T., Moldrup, P., Greve, M. H., et al. (2025). Physics-Informed Neural Networks for Estimating a Continuous Form of the Soil Water Retention Curve From Basic Soil Properties. Water Resources Research, 61(3), e2024WR038149. https://doi.org/10.1029/2024WR038149
