From hydrological laws to catchment models: A process-based model matches deep-learning accuracy on three continents

Xilin Xia

AI summary · arXiv: 2611.06382v1 · cs.LG

From hydrological laws to catchment models: A process-based model matches deep-learning accuracy on three continents

Xilin Xia

How to read this summary? Illustrated guide

The essentials

A low-dimensional, process-based catchment model matched or exceeded single LSTM models across British, US, and Australian catchments, suggesting shared catchment-scale laws of form.

Based on the abstract only: check the details in the original paper.

Question

The paper asks whether a low-dimensional catchment model constrained by physical laws can match the predictive accuracy of neural networks using the same inputs. It also examines whether the model’s constitutive forms can be shared across catchments while learning only catchment-specific parameters.

Method

The authors develop CaRT+ by iteratively refining a parsimonious model constrained by conservation, dissipation, and aggregation; CaRT+ has 14 states and 35 parameters per catchment and uses no daily neural network. They evaluate it on 602 British and 593 US gauged catchments in held-out years, and on 505 drier Australian gauged catchments not used in development, comparing it with stateful and entity-aware LSTMs.

Findings

Result In held-out years, single CaRT+ models achieved median NSEs of 0.863 in Britain and 0.765 in the US.
See the passage in the paper
On 602 British and 593 US gauged catchments, single CaRT+ models reach median Nash-Sutcliffe efficiencies of 0.863 and 0.765 in held-out years.
Result In Britain and the US, single CaRT+ models performed comparably to the two single LSTM models, while three-seed LSTM ensembles retained a small median paired NSE advantage.
See the passage in the paper
A stateful LSTM reaches 0.869 and 0.751 and an entity-aware LSTM 0.862 and 0.754, while three-seed LSTM ensembles remain ahead of the CaRT+ ensemble by 0.003-0.012 in median paired NSE.
Result On the Australian catchments, CaRT+ achieved a higher median NSE than either LSTM model.
See the passage in the paper
On 505 drier Australian gauged catchments unused in development, single CaRT+ models reach 0.778 against 0.762 and 0.753 for the two LSTMs.
Authors’ interpretation The authors interpret their results as evidence that catchment-scale laws can specify shared model forms, with catchment-specific parameters learned from data.
See the passage in the paper
The results suggest that catchment-scale laws exist as laws of form, with the constitutive families shared and only their parameters learned.

Limitations

The reported evaluation is limited to gauged catchments in Britain, the US, and Australia, so the results do not establish performance in other regions or settings.

Why it matters

The authors argue that physically constrained, low-dimensional models can attain deep-learning-level accuracy without a daily neural network. Their results support developing catchment models with shared constitutive families and catchment-specific learned parameters.

Written by AI from the paper. Passages are checked word for word against the text we analysed; the rest may contain errors.

AI-generated keywords: rainfall–runoff modelling; catchment hydrology; physically constrained models; neural networks; model generalization

Read further

Comprehensive Summary
Layman's Summary
Blog article
Original abstract

The paper asks whether a relatively simple, physically structured catchment model can match neural networks in rainfall–runoff prediction. It addresses the longstanding question of whether hydrological laws exist at the catchment scale. The authors propose that such laws need not define one detailed model. Instead, they can constrain a model’s structure while allowing its parameters to be learned for individual catchments. The authors identify three principles for constructing a lumped catchment model: conservation, dissipation, and aggregation. Conservation requires connected water stores and fluxes. Dissipation constrains the direction or sign of fluxes. Aggregation accounts for how unresolved catchment composition affects fluxes, including through effective relationships across threshold distributions and through memory of past forcing. Guided by these principles, the authors develop CaRT, a parsimonious model with 11 parameters per catchment. They then refine it iteratively, using AI agents to identify performance deficits and propose changes informed by hypotheses from the literature. The resulting model, CaRT+, has 14 states and 35 parameters per catchment and uses no daily neural network. According to the abstract, many refinements account for threshold aggregation or introduce variables that influence fluxes without storing water. The evaluation covers gauged catchments in Britain, the United States, and Australia. Across 602 British catchments, CaRT+ achieves a median Nash–Sutcliffe efficiency (NSE) of 0.863 in held-out years. This is close to the scores of the stateful LSTM (0.869) and entity-aware LSTM (0.862). Across 593 US catchments, CaRT+ scores 0.765, compared with 0.751 for the stateful LSTM and 0.754 for the entity-aware LSTM. However, three-seed LSTM ensembles outperform the CaRT+ ensemble by 0.003–0.012 in median paired NSE. Across 505 drier Australian catchments, which were not used during model development, CaRT+ scores 0.778, compared with 0.762 and 0.753 for the two LSTM approaches. These results suggest that the process-based model can match or outperform the reported single-network benchmarks in several settings, including a separate region with drier conditions. The authors conclude that catchment-scale hydrological laws may be laws of model form: model families can be shared across catchments while their parameters are learned separately. The Australian evaluation provides evidence that the approach may transfer beyond the development regions. The available material imposes important limits on these conclusions. Only the title and abstract are provided, so the data processing, train–test design, uncertainty analysis, baseline configurations, and statistical tests cannot be assessed. The abstract does not report results for the original 11-parameter CaRT model, quantify the contribution of individual refinements, or establish how broadly the findings generalize beyond the specified catchments and evaluation setup. It also reports a small advantage for LSTM ensembles in Britain and the United States. Therefore, the evidence does not show that CaRT+ is consistently more accurate than neural-network ensembles.

Can a simple catchment model keep up with deep learning?

Rainfall–runoff prediction is a central challenge in hydrology: given weather and catchment conditions, how much water will flow into rivers, and when? Deep-learning models can make accurate predictions, but their internal workings may be difficult to interpret. Process-based models, by contrast, represent water movement through stores and flows. The question is whether a model built around physical structure can achieve accuracy comparable to neural networks.

In “From hydrological laws to catchment models: A process-based model matches deep-learning accuracy on three continents,” Xilin Xia explores a nuanced idea: hydrological laws at the catchment scale may not prescribe one exact model. Instead, they may define principles that guide a shared model structure, while allowing its parameters to be learned separately for each catchment.

Three principles for model structure

The proposed framework rests on three principles. Together, they aim to preserve meaningful hydrological structure without requiring every catchment to be represented in exhaustive detail.

Conservation: Water stores and flows should be connected so that water entering, leaving, or remaining in the system is accounted for.
Dissipation: The direction or sign of flows should be constrained in physically sensible ways.
Aggregation: A catchment contains varied, unresolved features. The model must account for how that variation shapes effective flow relationships, including the effects of thresholds and the memory of past conditions.

These principles first guide CaRT, a relatively parsimonious model with 11 parameters per catchment. The authors then refine the model iteratively. AI agents identify places where performance falls short and propose changes informed by hypotheses from the hydrology literature. The resulting CaRT+ has 14 states and 35 parameters per catchment, and does not use a daily neural network.

According to the abstract, many refinements represent the combined effects of catchment features with different thresholds. Others add variables that affect flows without storing water themselves. The progression from CaRT to CaRT+ therefore illustrates how a physically structured model can become more flexible while retaining a process-based design.

How CaRT+ performed

The evaluation covers gauged catchments in Britain, the United States, and Australia. Performance is reported using the Nash–Sutcliffe efficiency (NSE), a measure of how well predictions reproduce observed variation. Higher values indicate better performance. The comparisons include stateful and entity-aware long short-term memory (LSTM) neural networks.

Britain: Across 602 catchments, CaRT+ reached a median held-out-year NSE of 0.863. The stateful LSTM scored 0.869, and the entity-aware LSTM scored 0.862.
United States: Across 593 catchments, CaRT+ scored 0.765, compared with 0.751 for the stateful LSTM and 0.754 for the entity-aware LSTM.
Australia: Across 505 drier catchments not used during model development, CaRT+ scored 0.778. The two LSTM approaches scored 0.762 and 0.753.

These results suggest that CaRT+ can match the reported single-network benchmarks in Britain and outperform them in the United States and Australia. The Australian result is particularly notable because those catchments were outside the regions used during development and were drier. It offers evidence that the approach may transfer to different conditions, though it does not by itself establish how broadly that transfer will hold.

What the results do—and do not—show

The comparisons are not uniformly in CaRT+’s favor. Three-seed LSTM ensembles outperform the CaRT+ ensemble by 0.003–0.012 in median paired NSE in Britain and the United States. The paper’s findings therefore do not show that CaRT+ is consistently more accurate than neural-network ensembles. A more measured conclusion is that a model built from physical structure can reach accuracy close to that of deep-learning approaches, and can outperform the reported single-network comparisons in some settings.

The abstract also leaves important questions unanswered. It does not provide enough detail to assess the data processing, train–test design, uncertainty analysis, baseline configurations, or statistical tests. Nor does it report the original CaRT model’s results or quantify how much each CaRT+ refinement contributes. Without those details, readers cannot tell which additions matter most or how robust the differences are.

A case for laws of model form

Xia’s central proposal is that catchment-scale hydrological laws may be laws of model form rather than one universal set of equations. On this view, catchments can share a structured model family while learning different parameters. CaRT+ makes that idea concrete: it combines constraints motivated by water movement with enough flexibility to represent catchment variation.

The results make a strong case that process-based modeling need not mean sacrificing predictive accuracy. They also offer promising evidence of performance beyond the regions used in development. But the abstract supports a carefully bounded claim—not that physical models have definitively surpassed deep learning, but that a structured catchment model can be a competitive alternative, with important questions about generalization and the role of individual refinements still open.

Created on 09 Nov. 2026

Was this synthesis useful?

Score: 0

Ask questions about this paper to our AI assistant

How to chat with this paper? Illustrated guide

Compare with the PDF

You can also chat with multiple papers at once here.

⚠The license of the paper does not allow us to build upon its content and the AI assistant only knows about the paper metadata rather than the full article.

AI assistant instructions?

Tools for this paper

Explain it differently 1 free rewrite left today

Save to a collection Group papers in collections to compare them, chat with them and export them. Continue with Google Create a free account Log in

Export BibTeX · RIS (Zotero, Mendeley, EndNote) · Obsidian (Markdown) · Notion (Markdown)

Advanced tools Premium

For researchers who follow a field closely: citation alerts, journal-club slides, version comparison and Zotero export. Discover Premium

Zotero

Citation alerts Citation alerts guide

Who works on it? The people and institutions who publish on its topic

Weekly letter Presented in "Artificial intelligence this week" of 9 November 2026. Artificial intelligence this week: When better scores miss the point

Journal club Premium+ Slides to present it Propose for the journal club

Versions What changed between versions

Talk about it in public Premium+ Press kit: release, questions, hype check Press kit guide

Papers around this one

See them on a map

Ranked by how close their title and abstract are to this paper’s. Select some to read them together.

Closest papers

Close Same question
Effective catchment hydrologic representations are low-dimensional but interaction-dependent
Yuan-Heng Wang, Hoshin V. Gupta · 2026 · Not summarized yet

This study tests how low-dimensional, mass-conserving catchment models gain predictive skill from additional states and interactions across US catchments.

Abstract
Hydrologic models must reconcile predictive flexibility with physical interpretability, yet the complexity required to reproduce catchment dynamics remains uncertain. Here we compare conceptual hydrologic models and neural networks built from mass-conserving computational units across 513 catchments in the contiguous United States, varying state dimensionality and cross-state information sharing. Under likelihood-based selection, median testing-period Kling-Gupta efficiency skill score (KGEss) increases from 0.83 with one allowable state to 0.89 with two, but reaches only 0.90 with five. KGE-based selection shows similarly rapid saturation. Cross-state information sharing provides an additio…
Close Another approach to the problem
Differentiable, Learnable, Regionalized Process‐Based Models With Multiphysical Outputs can Approach State‐Of‐The‐Art Hydrologic Prediction Accuracy
D. Feng, Jiang-Tao Liu, K. Lawson et al. · 2022 · Water Resources Research · cited 274 times · Not summarized yet

The δ models combine differentiable process-based structures with regionalized parameters to approach LSTM streamflow accuracy and predict additional water-cycle variables.

Abstract
Predictions of hydrologic variables across the entire water cycle have significant value for water resources management as well as downstream applications such as ecosystem and water quality modeling. Recently, purely data‐driven deep learning models like long short‐term memory (LSTM) showed seemingly insurmountable performance in modeling rainfall runoff and other geoscientific variables, yet they cannot predict untrained physical variables and remain challenging to interpret. Here, we show that differentiable, learnable, process‐based models (called δ models here) can approach the performance level of LSTM for the intensively observed variable (streamflow) with regionalized parameterizatio…
Close Another approach to the problem
Process-Aware AI for Rainfall-Runoff Modeling: A Mass-Conserving Neural Framework with Hydrological Process Constraints
M. Farmani, Hoshin V. Gupta, A. Behrangi et al. · 2026 · arXiv.org · preprint, not peer reviewed · Not summarized yet

The Mass-Conserving Perceptron incrementally adds hydrological process representations to a conservation-constrained neural framework for rainfall-runoff prediction.

Abstract
Machine learning models can achieve high predictive accuracy in hydrological applications but often lack physical interpretability. The Mass-Conserving Perceptron (MCP) provides a physics-aware artificial intelligence (AI) framework that enforces conservation principles while allowing hydrological process relationships to be learned from data. In this study, we investigate how progressively embedding physically meaningful representations of hydrological processes within a single MCP storage unit improves predictive skill and interpretability in rainfall-runoff modeling. Starting from a minimal MCP formulation, we sequentially introduce bounded soil storage, state-dependent conductivity, vari…
Close Broader context
When physics gets in the way: an entropy-based evaluation of conceptual constraints in hybrid hydrological models
Manuel Álvarez Chaves, Eduardo Acuña Espinoza, Uwe Ehret et al. · 2026 · Hydrology and Earth System Sciences · cited 2 times · Not summarized yet

This study evaluates how conceptual physical constraints affect hybrid hydrological model behavior using an entropy-based approach.

Abstract
Abstract. Merging physics-based with data-driven approaches in hybrid hydrological modeling offers new opportunities to enhance predictive accuracy while addressing challenges of model interpretability and fidelity. Traditional hydrological models, developed using physical principles, are easily interpretable but often limited by their rigidity and assumptions. In contrast, machine learning methods, such as Long Short-Term Memory (LSTM) networks, offer exceptional predictive performance but are often criticized for their black-box nature. Hybrid models aim to reconcile these approaches by imposing physics to constrain and understand what the ML part of the model does. This study introduces a…
Close Another approach to the problem
Distributed Hydrological Modeling With Physics‐Encoded Deep Learning: A General Framework and Its Application in the Amazon
Chao Wang, Shi-Jie Jiang, Yi Zheng et al. · 2024 · Water Resources Research · cited 103 times · Not summarized yet

This framework combines a distributed process-based model with neural parameter mapping and learned replacements for uncertain processes in the Amazon.

Abstract
While deep learning (DL) models exhibit superior simulation accuracy over traditional distributed hydrological models (DHMs), their main limitations lie in opacity and the absence of underlying physical mechanisms. The pursuit of synergies between DL and DHMs is an engaging research domain, yet a definitive roadmap remains elusive. In this study, a novel framework that seamlessly integrates a process‐based hydrological model encoded as a neural network (NN), an additional NN for mapping spatially distributed and physically meaningful parameters from watershed attributes, and NN‐based replacement models representing inadequately understood processes is developed. Multi‐source observations are…
Close Broader context
Hydrological Concept Formation inside Long Short-Term Memory (LSTM) networks
Thomas Lees, S. Reece, Frederik Kratzert et al. · 2021 · Hydrology and Earth System Sciences · cited 196 times · Not summarized yet

This study interprets LSTM internal states to examine whether their learned rainfall-runoff relationships represent hydrological concepts.

Abstract
Abstract. Neural networks have been shown to be extremely effective rainfall-runoff models, where the river discharge is predicted from meteorological inputs. However, the question remains, what have these models learned? Is it possible to extract information about the learned relationships that map inputs to outputs? And do these mappings represent known hydrological concepts? Small-scale experiments have demonstrated that the internal states of Long Short-Term Memory Networks (LSTMs), a particular neural network architecture predisposed to hydrological modelling, can be interpreted. By extracting the tensors which represent the learned translation from inputs (precipitation, temperature) t…
Close Same question
Rainfall-Runoff Simulation and Interpretation in Great Britain using LSTMs
Thomas Lees, M. Buechel, B. Anderson et al. · 2021 · Not summarized yet

This study evaluates LSTM and entity-aware LSTM rainfall-runoff predictions across British catchments and interprets their simulated hydrological processes.

Abstract
<p>Techniques from the field of machine learning have shown considerable promise in rainfall-runoff modelling. This research offers three novel contributions to the advancement of this field: a study of the performance of LSTM based models in a GB hydrological context; a diagnosis of hydrological processes that data-driven models simulate well but conceptual models struggle with; and finally an exploration of methods for interpreting the internal cell states of the LSTMs. </p><p>In this study we train two deep learning models, a Long Short Term Memory (LSTM) Network and an Entity Aware LSTM (EALSTM), to simulate discharge for 518 catchments across Great Britain using a newly published d…
Close Same method, another problem
The suitability of differentiable, physics-informed machine learning hydrologic models for ungauged regions and climate change impact assessment
D. Feng, H. Beck, K. Lawson et al. · 2023 · Hydrology and Earth System Sciences · cited 178 times · Not summarized yet

This study uses regionalized differentiable process-based models to test streamflow predictions in ungauged regions and under climate change.

Abstract
Abstract. As a genre of physics-informed machine learning, differentiable process-based hydrologic models (abbreviated as δ or delta models) with regionalized deep-network-based parameterization pipelines were recently shown to provide daily streamflow prediction performance closely approaching that of state-of-the-art long short-term memory (LSTM) deep networks. Meanwhile, δ models provide a full suite of diagnostic physical variables and guaranteed mass conservation. Here, we ran experiments to test (1) their ability to extrapolate to regions far from streamflow gauges and (2) their ability to make credible predictions of long-term (decadal-scale) change trends. We evaluated the models bas…

Go further

How are these papers found?

We compare the title and abstract of this paper with those of every paper summarized on this site, and with the papers Semantic Scholar recommends, those it cites and those citing it (for a much-cited paper, its most recent citations and, from OpenAlex, the most cited ones). Among the prior work and the follow-ups, the papers that use each other in depth, a method it uses and the much-cited follow-ups come a little higher. The levels (very close, close, related) come from the similarity of the texts, not from a judgment of quality; * marks a paper without an abstract, compared by its title. The sentence under a paper is written by an AI from the two abstracts: check it in the papers. A paper marked not relevant by three readers leaves the list.

Updated on 9 November 2026.

Disclaimer: The AI-based summarization tool and virtual assistant provided on this website may not always provide accurate and complete summaries or responses. We encourage you to carefully review and evaluate the generated content to ensure its quality and relevance to your needs.

From hydrological laws to catchment models: A process-based model matches deep-learning accuracy on three continents

The essentials

Question

Method

Findings

Limitations

Why it matters

Read further

Definitions

Can a simple catchment model keep up with deep learning?

Three principles for model structure

How CaRT+ performed

What the results do—and do not—show

A case for laws of model form

Ask questions about this paper to our AI assistant

Tools for this paper

Advanced tools Premium

Papers around this one

Closest papers

Go further

Cite this paper (BibTeX)