Technical Summary: From Conceptual Hydrologic Models to Conceptually Interpretable Neural Networks

Problem Statement

The hydrologic community faces a persistent dichotomy between physical-conceptual (PC) models and data-driven machine learning (ML) approaches. PC models offer interpretability through explicit process representations but often rely on overly compressed state-space assumptions. Conversely, flexible black-box ML models, such as Long Short-Term Memory (LSTM) networks, achieve high predictive accuracy across diverse conditions but lack internal representations that translate into scientific understanding. This gap is particularly acute in snow-dominated regions, where representing coupled snow–water dynamics requires accounting for complex phase transitions and storage mechanisms. While recent efforts have integrated physical constraints (e.g., mass conservation) into ML architectures, there remains a need for frameworks that bridge the gap between prescribed process pathways and flexible functional approximations without sacrificing conceptual interpretability.

Methodology

This study advances the Mass-Conserving Perceptron (MCP) framework, a recurrent neural network unit that explicitly enforces mass conservation at the nodal level. The research transitions from viewing conceptual models and neural networks as distinct paradigms to treating them as alternative representations of the same precipitation–storage–runoff (PSR) problem.

Core Computational Units:
The study utilizes three primary MCP-based computational units:

1. HMCP(1, AL): A single-node unit representing soil-moisture-like storage, with output and loss gates dependent on storage state and potential evapotranspiration.
2. SNOWMCP(1, BQ): A single-node unit representing snow water equivalent (SWE), featuring rain–snow partitioning, melting, and loss gates dependent on temperature and internal SWE state.
3. HMCP(1, ALMR): An HMCP variant augmented with a mass-relaxation gate to represent unobserved mass exchange with the environment.

Experimental Framework:
The authors evaluated these units across 513 minimally disturbed catchments in the conterminous United States (CAMELS-US) using data from water years 1982–2008. The experimental design progressed through several architectural hypotheses:

• Coupled Hydrologic Models (HYDROMCP): Two-node architectures coupling HMCP and SNOWMCP in serial (routing) or parallel (bypass) configurations. These models tested explicit process hypotheses regarding whether snowmelt routes through soil storage or bypasses it.
• Cross-Node State Sharing: Variants of HYDROMCP where gating functions in one node access the cell state of the other (e.g., soil moisture influencing snowmelt gates), testing the hypothesis of coupled state interactions.
• Functional Approximation Networks: Single-layer, feedforward neural networks constructed from parallel MCP units. These include:
  • Distributed-State (DS) Networks: Homogeneous networks of 1–5 HMCP or SNOWMCP nodes.
  • Information Sharing: Variants with full state sharing (SALO), partial sharing (adjacent nodes only), and multi-output layers.
  • Mixed-Type Networks: Architectures combining HMCP and SNOWMCP nodes.
• Benchmarks: Performance was compared against standard LSTM networks and selected using two criteria: the Kling-Gupta Efficiency ($KGE^{ss}$) for predictive skill and the Akaike Information Criterion (AIC) for balancing skill with model parsimony (parameter count and state dimensionality).

Key Contributions

1. Unified Representation: The paper establishes a modeling paradigm where conceptual hydrologic models are reformulated as physically constrained, conceptually interpretable neural networks. It demonstrates that the distinction between PC models and neural networks often lies in architectural structure rather than fundamental philosophy.
2. Snow-Water MCP Framework: A specific extension of the MCP framework to snow-dominated environments, introducing the SNOWMCP unit and evaluating its coupling with soil-moisture representations.
3. Complexity Analysis: A systematic evaluation of the minimum model complexity required for different catchments. The study quantifies the trade-offs between the number of storage states, the type of computational units, and the necessity of cross-node information sharing.
4. Model Selection Guidance: The development of an AIC- and KGE-based selection framework that identifies basin-specific optimal architectures, balancing predictive accuracy with structural parsimony.

Results

Performance and Complexity:

• State Dimensionality: Across CONUS, the median $KGE^{ss}$ increased from 0.82 for one-state networks to 0.89 for two-state networks and 0.90 for five-state networks. The study found diminishing aggregate gains beyond two states, suggesting that two states often suffice for robust representation.
• Snow vs. Soil: In snow-dominated basins, coupling snow and soil components (HYDROMCP) significantly improved performance over single-state models. The median $KGE^{ss}$ for snowy basins increased from 0.72 (single-state HMCP) to 0.89 (two-state HYDROMCP).
• Information Sharing: Cross-node state sharing (SALO) provided substantial benefits, particularly in high-elevation, snow-influenced regions. For example, in the Rocky Mountains, the median $KGE^{ss}$ for a three-node SNOWMCP network increased from 0.90 (no sharing) to 0.96 (with sharing).
• Functional Approximation: The "functional approximation" network formulation (where HMCP and SNOWMCP receive total precipitation and outputs are combined via Softmax) achieved performance comparable to the more rigid "hydrologic model" formulations (serial/parallel routing) but with greater representational flexibility.

Comparison with LSTMs:

• Predictive Skill: Basin-specific best MCP networks achieved a median $KGE^{ss}$ of 0.90, comparable to the best-performing LSTM networks.
• Parameter Efficiency: MCP-based networks achieved this performance with significantly fewer parameters on average (69.45 for best MCP vs. 115.86 for best LSTM).
• Selection Criteria: Under AIC-based selection (favoring parsimony), MCP-based networks were preferred in 63% of basins, whereas under KGE-based selection (favoring pure skill), the preference was nearly split (49.3% MCP vs. 48.9% LSTM).

Regional Variations:

• Snow Regimes: Performance gains from snow-informed representations and state sharing were most pronounced in high-elevation western basins (e.g., Rocky Mountains, Sierra Nevada). In these regions, models with explicit snow states and cross-node sharing outperformed single-state models by large margins.
• Non-Snowy Basins: In non-snowy basins, multi-state architectures still provided benefits, suggesting that multiple flow pathways are often necessary even in the absence of dominant snow processes.

Significance and Claims

The authors claim that this study provides an empirical basis for identifying the numbers, types, and interactions of states needed for hydrologic representation. By treating conceptual models and neural networks as a continuum of representations, the framework allows for the systematic testing of hydrologic hypotheses (e.g., the necessity of snow-soil coupling) within a data-driven, mass-conserving structure.

The paper posits that:

1. Interpretability is Retained: MCP-based networks retain conceptual interpretability because their internal states and gating mechanisms correspond to physical concepts (storage, fluxes, phase changes), unlike black-box LSTMs.
2. Parsimony is Achievable: Compact, basin-specific architectures (often two-state) can recover most of the predictive skill of more complex models, offering a practical path toward "models of everywhere" that are both accurate and scientifically grounded.
3. Future Directions: The authors suggest that future work should test joint training against multiple hydrologic responses (e.g., streamflow, SWE, groundwater) to better constrain the learned internal states and further strengthen the link between statistical learning and physical reasoning. They also highlight the potential for using these frameworks to generate compositional hypotheses and explore latent hypothesis spaces via generative modeling.

The study concludes that the MCP framework successfully bridges physical-conceptual reasoning with modern data-driven modeling, offering a transparent and scientifically testable approach to catchment-scale PSR dynamics.