Wang et al. (2025) Using Machine Learning to Discover Parsimonious and Physically‐Interpretable Representations of Catchment‐Scale Rainfall‐Runoff Dynamics

⚠️ Warning: This summary was generated from the abstract only, as the full text was not available.

Identification

• Journal: Water Resources Research
• Year: 2025
• Date: 2025-12-01
• Authors: Yuan‐Heng Wang, H. Gupta
• DOI: 10.1029/2025wr040178

Research Groups

Not specified in the abstract.

Short Summary

This paper explores the development of physically interpretable machine learning models for dynamical systems, demonstrating that Mass-Conserving-Perceptron (MCP) based networks with a distributed-state mechanism can achieve both physical interpretability and good predictive performance in catchment-scale streamflow modeling with minimal complexity.

Objective

• To develop parsimonious, minimally-optimal, and physically-interpretable machine learning models for dynamical systems, specifically for streamflow modeling, that overcome the interpretability challenges of traditional ML while maintaining good predictive performance.

Study Configuration

• Spatial Scale: Spatially-lumped catchment-scale modeling.
• Temporal Scale: Dynamical systems, implying continuous or discrete time steps relevant to hydrological processes.

Methodology and Data

• Models used: Mass-Conserving-Perceptron (MCP) based network architectures; "distributed-state" network with context-dependent gating and "information-sharing".
• Data sources: Not specified in the abstract, but implies hydrological data for streamflow modeling.

Main Results

• Both physical interpretability and good predictive performance can be achieved using a "distributed-state" network with context-dependent gating and "information-sharing" across nodes.
• The distributed-state mechanism ensures a sufficient number of temporally-evolving properties of system storage.
• Information-sharing ensures proper synchronization of such properties.
• MCP-based ML models with only a few layers (up to two) and relatively few physical flow pathways (up to three) can play a significant role in ML-based streamflow modeling.

Contributions

• Proposes a novel approach using Mass-Conserving-Perceptron (MCP) based network architectures to build inherently physically-interpretable machine learning models for dynamical systems.
• Demonstrates that parsimonious ML models (few layers, few pathways) can achieve both high predictive performance and physical interpretability in catchment-scale streamflow modeling.
• Introduces "distributed-state" and "information-sharing" mechanisms to enhance interpretability and synchronization in ML hydrological models.

Funding

Not specified in the abstract.

Citation

@article{Wang2025Using,
  author = {Wang, Yuan‐Heng and Gupta, H.},
  title = {Using Machine Learning to Discover Parsimonious and Physically‐Interpretable Representations of Catchment‐Scale Rainfall‐Runoff Dynamics},
  journal = {Water Resources Research},
  year = {2025},
  doi = {10.1029/2025wr040178},
  url = {https://doi.org/10.1029/2025wr040178}
}