Graph-based Automated Discovery of Concise Soil Hydraulic Functions from Data: Beyond Empirical Formulations

arXiv:2605.19379 · physics.flu-dyn, cs.SC, physics.geo-ph · Submitted 2026-05-19 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.

Kai: We're looking at the paper: "Graph-based Automated Discovery of Concise Soil Hydraulic Functions from Data".

Mira: Here's the gist of the paper. Soil hydraulic functions are fundamental to modelling water flow and transport in vadose-zone hydrology and are central to a wide range of hydrological and geoscientific applications Here we present a graph-based automated model discovery framework for discovering explicit soil hydraulic functions directly from experimental data Across 249 real soil samples spanning diverse textural classes, the discovered functions achieve more accurate predictions of unsaturated hydraulic conductivity than the MvG model This work demonstrates that data-driven model discovery can move beyond traditional empirical derivation and provide a promising route for developing accurate and explicit constitutive models.

Kai: Automated Model Discovery Framework The proposed framework encodes an equation as a directed acyclic graph G=(V,E), where nodes correspond to variables or operators and edges capture computational dependencies To capture variability across different soil samples, edge features encode parametric dependencies through fixed constants or undetermined parameters.

Mira: The discovery process follows a “generation-evaluation-evolution” paradigm implemented through genetic programming.

Kai: This involves randomly generating candidate graphs, applying crossover and mutation operations to explore the expression space, and evaluating each graph using a loss function.

Paper summary: Kai: So we're looking at this paper today called "Graph-based Automated Discovery of Concise Soil Hydraulic Functions from Data: Beyond Empirical Formulations". Mira, what’s the main idea here?

Mira: Basically, they’re tackling how soil hydraulic functions are modeled. right now, most people use these expert-designed empirical formulas like the Mualem van Genuchten model. but those models have trouble being both accurate and compact when you look at different soil types. this paper proposes a graph-based method to discover these functions directly from experimental data instead of just guessing them based on existing formulas.

Lev: So it’s about building a system that finds the actual equation, not just fitting some curve to the data points we already have. that sounds like it could be useful for checking if the underlying physics are actually what we think they are or if we're missing something fundamental.

Kai: Exactly, and they claim this discovery works across two hundred forty-nine real soil samples that cover a lot of different textures <ref:2605.19379#pg2,across 249 real soil samples>. the core thesis is that this automated discovery method finds a specific mathematical structure for water retention and hydraulic conductivity that is different from the classic MvG model.

Mira: They show that this newly discovered structure leads to better predictions for unsaturated hydraulic conductivity compared to the MvG model, achieving an overall root mean square error of two point five zero two for log of K r compared to three point zero six nine for the MvG model, as stated in the abstract >

Lev: That difference in error numbers is significant because it means the new form is statistically better at predicting how water moves through the soil than what we’re currently using for many applications. but we need to remember that this discovery relies on finding a structure that's inherently analytically integrable, so it can actually be used to calculate the hydraulic conductivity from an expression.

Kai: That’s a key part of why they focus on the analytical structure, because if you can integrate it, you get a closed-form equation for the soil hydraulic function instead of just a black box prediction. but what about how they actually do this discovery?

Mira: They use something called an automated model discovery framework based on encoding an equation as a directed acyclic graph. so they are defining variables and operators as nodes, and the way those variables connect is captured by edges that describe dependencies, including parameters or constants.

Lev: That sounds like genetic programming is involved, right? exploring the space of possible equations randomly by generating candidate graphs and then using crossover and mutation to refine them based on a loss function. that’s how they search for the best structure >

Kai: Right, so they are trying out millions of potential equation shapes against their soil data. the loss function is what guides it, composed of two parts: data loss, which is like mean squared error, and a complexity penalty based on how many soil specific parameters are in the graph.

Paper summary: Mira: That complexity penalty is interesting because it stops them from just finding an infinitely complicated equation that fits noise perfectly; they want something relatively simple. and they found that the soil-specific parameters are tied to things like particle size, where coarse soils favor one parameter value and fine soils favor another >

Lev: So the framework isn't just blindly searching for *any* fit; it’s searching for a compact, physically relevant structure that has a specific relationship with the soil texture properties. that makes it more interpretable than some of those deep learning approaches I mentioned earlier, where you get a prediction but don't really know what the equation means.

Kai: That makes sense because we’re trying to move away from those black box predictors where you can't see the mechanism. they are aiming for something that is both accurate and analytically sound, which is a big thing for real world modeling >

Mira: And it seems they succeeded in finding a tangent-based form that has comparable water retention fitting but provides much improved hydraulic conductivity prediction, as shown by that RMSE of two point five zero two versus the MvG model's three point zero six nine >

Lev: That’s good performance for a model derived purely from data without starting with the MvG assumptions baked in. but we should also look at what they say about generalizability; how well this discovered function works outside of those two hundred forty-nine specific soil samples > <ref:2605.19379#pg2>

Kai: They tested it on the Unsaturated Soil Hydraulic Database, UNSODA, which is another set of data covering different regions and soil types. the results suggest that the discovered model has superior distributional robustness because it effectively eliminates that long error tail when predicting the hydraulic conductivity function >

Mira: It’s interesting how this data-driven discovery process led to a functional form that is fundamentally distinct from what we see in classical soil physics literature, even though it still retains an analytical structure for integration >

Lev: The fact that they retained an analytical structure is important because it means the output isn't just a curve you have to use for look-up tables; it's a formula you can plug things into, which is what engineers and hydrologists need >

Kai: So, to sum up this paper called "Graph-based Automated Discovery of Concise Soil Hydraulic Functions from Data: Beyond Empirical Formulations", they used an automated graph framework to discover a new soil hydraulic function that outperforms the MvG model in predicting hydraulic conductivity across diverse soils, and that new function is analytically integrable >

Mira: And it matters because it provides a path toward developing explicit constitutive models that are both accurate and robust, moving away from purely empirical formulations where the underlying physics isn't fully captured >

Lev: For practical use on real hardware or large-scale simulations, having an equation you can integrate is much more reliable than a complex neural network that you can't easily debug when it starts failing in an unexpected scenario >

Conclusion: Kai: So, this paper is about taking experimental soil data and using a graph to automatically discover the actual mathematical formula for how water moves through that soil.

Mira: Yeah, they're showing how you can move past just fitting curves to find a real physical structure that describes the soil hydraulic functions.

Lev: From an engineering standpoint, I’m really interested in whether this discovered structure is something we can actually build on real hardware or simulate reliably.

Kai: They tested this against two hundred forty-nine samples from a huge database, and the result is that their new formula for how water flows is more accurate than the standard Mualem van Genuchten model.

Mira: That's significant because it means we have a better tool to predict unsaturated hydraulic conductivity when we don't know the exact physics beforehand.

Lev: A better prediction is good, but what they found about the parameters for soil texture—that coarse soils lean toward one parameter and fine soils toward another—is something that would need careful testing on real-world data.

Kai: So, the authors managed to find a form that has both high predictive power and stays relatively simple enough to be analytically integrated into an actual equation.

Mira: It's about balancing having a highly accurate result with keeping the underlying mathematical structure manageable for modeling.

Lev: It's interesting because they found this new functional form is fundamentally different from what we usually see in soil physics literature, even though it still has that integrable structure.

Kai: And that analytical structure is what makes it useful; it lets you derive the hydraulic conductivity directly from an expression rather than just using a lookup table.

Mira: It provides a new pathway for developing explicit constitutive models, which means we can actually write down the math describing the soil behavior instead of relying on black box predictions.

Lev: If we can get that kind of explicit model, it makes designing better drainage systems or predicting water movement in complex subsurface environments much more feasible.

Hao Xu, Jinshen Sun, Yuntian Chen, *Dongxiao Zhang

Zhejiang Key Laboratory of Industrial Intelligence and Digital Twin · Department of Electrical Engineering, Tsinghua University · CNPC Engineering Technology R&D Company Limited

physics.flu-dyn, cs.SC, physics.geo-ph

Submitted: 2026-05-19

Updated: 2026-11-04

Code: https://github.com/woshixuhao/Discovery_of_unsaturated_soil_model

License: http://creativecommons.org/licenses/by-nc-nd/4.0/

Importance score: 89/100

The gist: Soil hydraulic functions are fundamental to modelling water flow and transport in vadose-zone hydrology and are central to a wide range of hydrological and geoscientific applications Here we present…

Key concepts

Graph Encoding
The framework represents a mathematical equation as a directed acyclic graph (DAG). Nodes represent variables or operators in the equation, and edges show how these components are computationally connected. Edge features encode parameters that can vary between different soil samples, allowing the model to capture sample-specific behavior.
Generation-Evaluation-Evolution
This is the core discovery process driven by genetic programming. The system randomly generates candidate equations (graphs), evaluates them using a loss function that penalizes both prediction errors and excessive complexity (number of parameters), and then uses crossover and mutation to evolve better models over many iterations.
Loss Function
The evaluation metric used to judge a candidate model consists of two parts: data loss, which measures the average squared error between the model's predictions and the experimental data, and a complexity penalty. This penalty discourages overly complex equations that might just memorize the data without generalizing well.
Soil Texture Dependence
Statistical analysis showed that soil-specific parameters are strongly related to soil texture. For instance, coarse soils favored one parameter setting (p=1), while fine soils favored another (p=3). This indicates that the model's parameters naturally adapt based on the physical properties of the soil particles.

Terminology

Summary

Soil hydraulic functions are fundamental to modelling water flow and transport in vadose-zone hydrology and are central to a wide range of hydrological and geoscientific applications Here we present a graph-based automated model discovery framework for discovering explicit soil hydraulic functions directly from experimental data Across 249 real soil samples spanning diverse textural classes, the discovered functions achieve more accurate predictions of unsaturated hydraulic conductivity than the MvG model This work demonstrates that data-driven model discovery can move beyond traditional empirical derivation and provide a promising route for developing accurate and explicit constitutive models. p. 1

Where the paper says it

This work demonstrates that data-driven model discovery can move beyond traditional empirical derivation and provide a promising route for developing accurate and explicit constitutive models. INTRODUCTION Soil hydraulic functions, most notably the soil-water retention curve (SWRC) and the hydraulic conductivity function (HCF), are indispensable for describing water flow in variably saturated porous media and therefore underpin a wide range of applications in hydrology, agriculture, and geoscience [1].

Automated Model Discovery Framework

The proposed framework encodes an equation as a directed acyclic graph G=(V,E), where nodes correspond to variables or operators and edges capture computational dependencies To capture variability across different soil samples, edge features encode parametric dependencies through fixed constants or undetermined parameters. The discovery process follows a “generation-evaluation-evolution” paradigm implemented through genetic programming. p. 4

Where the paper says it

This design enables the discovery of equations that incorporate soil-specific parameters from observation data. The discovery process follows a “generation-evaluation-evolution” paradigm implemented through genetic programming.

This involves randomly generating candidate graphs, applying crossover and mutation operations to explore the expression space, and evaluating each graph using a loss function. p. 4
Where the paper says it

Following graph initialization, evolutionary operations, such as crossover and mutation, are applied to explore new candidate graphs. In the crossover step, two graphs exchange randomly selected subgraphs.

The loss function is composed of two terms: data loss (average mean squared error) and a complexity penalty proportional to the number of soil-specific parameters. p. 5
Where the paper says it

C i denotes the optimized values of the undetermined parameters for case i, and ε is the hyperparameter controlling the complexity penalty associated with the undetermined parameters. The loss function is composed of two terms: data loss, which is calculated as the average mean squared error of the candidate equation across all cases, and a complexity penalty proportional to the number of soil-specific parameters.

Discovery and Validation Datasets

The equation discovery is based on classic soil water retention curves and hydraulic conductivity measurements originally compiled for the Mualem–van Genuchten (MvG) model. The framework is further tested on 249 cases from the Unsaturated Soil Hydraulic Database (UNSODA) [25] to evaluate generalizability across diverse soil types and geographic regions. p. 3

Where the paper says it

We further evaluate the discovered functions on 249 soil samples from the Unsaturated Soil Hydraulic Database (UNSODA) [25], which includes field and laboratory measurements spanning diverse soil types and geographic regions. Across these soils, the discovered model achieves consistently strong predictive performance.

Model Performance and Comparison

The discovered model yields a fundamentally different tangent-based form with comparable water retention fitting but improved hydraulic conductivity prediction compared to the MvG model. The modified model achieves an overall RMSE of 2. p. 3

Where the paper says it

We further evaluate the discovered functions on 249 soil samples from the Unsaturated Soil Hydraulic Database (UNSODA) [25], which includes field and laboratory measurements spanning diverse soil types and geographic regions. Across these soils, the discovered model achieves consistently strong predictive performance.

502 for log(Kr), which is better than 3. p. 7
Where the paper says it

To enhance numerical stability and facilitate comparisons across soil samples, the targeted variable is set to (1/h)×100. For the graph-based representation, the operator set to generate potential functions comprised add, mul, exp, pow, log, sin, cos, and tan.

069 for the classic MvG model. p. 5
Where the paper says it

More details can be found in our previous work [27]. Datasets for equation discovery and validation The equation discovery in this study is based on a set of classic soil water retention curves and hydraulic conductivity measurements originally compiled and used in the development of the Mualem–van Genuchten (MvG) model [4,18].

Furthermore, the modified model demonstrates superior distributional robustness by effectively eliminating the long-error tail in the prediction of HCF. p. 1
Where the paper says it

This work demonstrates that data-driven model discovery can move beyond traditional empirical derivation and provide a promising route for developing accurate and explicit constitutive models. INTRODUCTION Soil hydraulic functions, most notably the soil-water retention curve (SWRC) and the hydraulic conductivity function (HCF), are indispensable for describing water flow in variably saturated porous media and therefore underpin a wide range of applications in hydrology, agriculture, and geoscience [1].

Parameter Relevance and Generalization

Statistical analysis revealed that soil-specific parameters are relevant to soil textures and properties. The choice of parameter p is strongly associated with soil particle size, with coarse-textured soils favoring p=1 and fine-textured soils favoring p=3. p. 8

Where the paper says it

Establishment of hydraulic conductivity function based on the discovered model Upon the automated identification of the SWRC function, we subsequently derive a physically consistent expression for the unsaturated hydraulic conductivity function. The generalized formulation between the relative hydraulic conductivity Kr and the SWRC can be defined by the following integral transform: 2 1/ 2 ( ) 1 ( ) 1       =        up low low dx h x dx h x Kr (6) This derivation follows the Mualem-type capillary framework and is therefore intended for soils in which hydraulic conductivity is primarily governed by capillary pore-size effects.

Both log(A) and β exhibit distinct spatial patterns when mapped to the soil texture triangle, suggesting they are influenced by soil composition. p. 7
Where the paper says it

To enhance numerical stability and facilitate comparisons across soil samples, the targeted variable is set to (1/h)×100. For the graph-based representation, the operator set to generate potential functions comprised add, mul, exp, pow, log, sin, cos, and tan.

Conclusion on Model Superiority

The proposed framework discovers a previously unexplored functional form for SWRC and its corresponding HCF that is distinct from the MvG model. The discovered model retains an analytical structure, allowing an explicit expression for hydraulic conductivity to be obtained through integration. p. 9

Where the paper says it

Notably, obtaining an explicit hydraulic conductivity function is intrinsically difficult, because conductivity depends on pore structure, connectivity and tortuosity, which are only partially encoded in water retention behaviour. Simple SWRC forms can preserve analytical tractability, but often with limited predictive fidelity.

This approach balances structural parsimony with predictive accuracy by identifying compact, accurate and analytically integrable expressions. p. 9
Where the paper says it

As demonstrated in Supplementary Materials S1, the exact analytical derivatives and constitutive sensitivities for both θ and Kr can be derived in compact, closed-form expressions. This structural simplicity renders the discovered model more amenable for deriving analytical solutions and for numerical and stochastic modelling without sacrificing predictive accuracy.

The work introduces a promising new functional form for hydraulic conductivity modelling and provides a pathway for interpretable, data-driven constitutive modelling. p. 1
Where the paper says it

This work demonstrates that data-driven model discovery can move beyond traditional empirical derivation and provide a promising route for developing accurate and explicit constitutive models. INTRODUCTION Soil hydraulic functions, most notably the soil-water retention curve (SWRC) and the hydraulic conductivity function (HCF), are indispensable for describing water flow in variably saturated porous media and therefore underpin a wide range of applications in hydrology, agriculture, and geoscience [1].

Supplementary Information

The analytical derivation of constitutive derivatives and parameter sensitivities is a fundamental requirement for robust hydrological modelling across scales The three capacity coefficients required for stochastic linearization are therefore Ch=(θs-θr)1/βAh 2(1+ν 2), CA=(θs-θr)1/βA 2h(1+ν 2), Cβ=-(θs-θr)Θ+1/β, as shown in the text. The final comparison shows that the modified model uses the fewest parameters and retains a compact analytical form while achieving the best overall prediction of Kr. p. 9

Where the paper says it

Moreover, the parameters A and β are structurally decoupled from the exponent system, avoiding the nested exponent terms in MvG model. As demonstrated in Supplementary Materials S1, the exact analytical derivatives and constitutive sensitivities for both θ and Kr can be derived in compact, closed-form expressions.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Automated model discovery framework: The graph-based representation allows for a broader exploration of the symbolic search space than human-driven trial-and-error, enabling the identification of compact, accurate and analytically integrable expressions. This means AI systems can discover new constitutive laws directly from experimental data without relying on preconceived functional forms like Mualem–van Genuchten.

  2. Explicit hydraulic conductivity derivation: The framework yields a closed-form expression for hydraulic conductivity through an integral transform, enabling the prediction of Kr accurately over a wide range of pressure heads, which is superior to models where conductivity depends on numerical evaluation.

  3. Enhanced generalization through parameter modification: The introduction of the modified model, such as Eq. (9), allows the AI system to handle diverse soil textures by introducing a discrete degree of freedom 'p' that accommodates a broader range of curve shapes without sacrificing compactness or interpretability. This improves predictive performance across the UNSODA database compared to classical models.

  4. Physical relevance and parameter correlation: Statistical analysis reveals that soil-specific parameters are strongly influenced by soil composition, providing a pathway for AI systems to discover constitutive relations where parameters have clear relationships with physical descriptors like characteristic pore radius r75 and maximum drainage slope.

Abstract

Soil hydraulic functions underpin water flow and transport modeling in the vadose zone, yet they are commonly represented by expert-designed formulations such as the Mualem-van Genuchten (MvG) model. Because these formulations rest on predefined functional assumptions, balancing predictive accuracy and parametric compactness across soil textures remains challenging. Here we introduce a graph-based equation discovery framework that distills soil hydraulic functions from experimental data. Reapplied to the original datasets that motivated the MvG parameterization, the framework recovers a concise water retention function together with an analytical unsaturated hydraulic conductivity expression whose mathematical structure departs from classical empirical forms. Evaluated across soils of contrasting texture, the discovered model predicts unsaturated hydraulic conductivity more accurately than MvG and its common variants, while its fitted parameters retain interpretable correlations with measurable soil physical properties. These findings suggest that data-driven discovery can yield compact, analytically tractable constitutive relations that move beyond fixed empirical formulations.

Sources

Related papers