Mass-Conserving Perceptron in Geoscientific Modeling

• Mass-Conserving Perceptron is a novel computational unit that enforces mass conservation and represents physical processes in a geoscientific context.
• It bridges traditional physical-conceptual models and machine learning by using trainable, gated recurrent architectures to partition mass fluxes.
• Empirical results demonstrate its capacity for hypothesis testing and improved predictive skill in hydrological applications across varying catchments.

Searching arXiv for papers on the Mass-Conserving Perceptron to ground the article in the cited literature. The Mass-Conserving Perceptron (MCP) is a physically interpretable computational unit for machine-learning-based modeling of geoscientific systems. It was introduced to bridge traditional physical-conceptual (PC) models, which encode conservation and process meaning, and ML models, especially gated recurrent neural network (GRNN) approaches, which can be much more accurate but are difficult to interpret physically. The MCP exploits the inherent isomorphism between the directed graph structures underlying both PC models and GRNNs to represent mass-conserving physical processes explicitly while learning process functions directly from data using off-the-shelf ML technology (Wang et al., 2023).

1. Origins and scientific motivation

The MCP emerged from a specific tension in geoscientific modeling. Decades of effort had been devoted to building PC models for predicting the time-series evolution of geoscientific systems, while more recent work showed that ML-based GRNN technology could produce much higher predictive accuracy. The difficulty was that improved prediction did not by itself yield improved physical understanding. The MCP was proposed as a way to combine physical constraints, especially mass conservation, with data-driven learning in a form that remains scientifically interpretable (Wang et al., 2023).

In its original formulation, the MCP is a recurrent (dynamical) node representing the input-state-output evolution of a control volume in a mass-conserving system. In hydrological examples, that control volume can represent soil moisture storage in a rainfall-runoff model. The fundamental quantities are the internal mass $X_t$, mass input $U_t$, mass output $O_t$, and unobserved loss $L_t$. This choice of variables makes the computational structure directly analogous to a conceptual storage-and-flux model rather than to an opaque latent-state RNN (Wang et al., 2023).

A recurring theme in the subsequent literature is that the MCP is intended not merely as a predictive model, but as a framework for scientific hypothesis testing, process diagnosis, and parsimonious representation discovery. This suggests that its significance lies as much in model structure as in forecast skill.

2. Formalism and mass-conservation mechanism

The core discrete-time balance equation is

$X_{t+1} = X_t + U_t - O_t - L_t .$

This is the defining conservation statement of the MCP. In gate-based form, the same mechanism is written as

$$X_{t+1} = (1 - G_O - G_L)\cdot X_t + G_U \cdot U_t,$$

with output and loss fluxes parameterized by

$O_t = G_O \cdot X_t,\qquad L_t = G_L \cdot X_t.$

The quantity $GR = 1 - G_O - G_L$ is the remember gate, and the normalization condition

$G_O + G_L + GR = 1$

enforces mass conservation at each timestep and node (Wang et al., 2023).

The gates are context-dependent and trainable. They can be parameterized as simple sigmoid functions or as arbitrary ML functions. A representative form is

$G_O = \sigma(a_O + b_O X_t)\cdot K_O,$

with analogous forms for other gates. For multiple gates, a Softmax normalization is used:

$U_t$0

This ensures positivity and normalization across competing pathways. The framework therefore uses bounded differentiable gates to represent physically interpretable process partitioning rather than abstract activation patterns (Wang et al., 2023).

The hydrological extensions add further physically meaningful gates. These include an input bypass mechanism with $U_t$1, a mass relaxation gate $U_t$2 for exchange of mass with the environment, and an input bias correction gate for learning systematic observational bias in inputs. Later formulations also impose additional constraints such as $U_t$3 for evapotranspiration. These additions preserve the basic principle that conservation is enforced by construction rather than imposed as a soft penalty (Wang et al., 2024).

3. Relation to physical-conceptual models and recurrent neural networks

The MCP is explicitly designed around a graph interpretation. Its architecture mimics the node-link structure of classical hydrological models, with each node embodying a conservation law and each link representing a process pathway. In this sense, the MCP uses a directly interpretable graph-based architecture while retaining compatibility with standard differentiable ML workflows such as PyTorch implementation and gradient-based training (Wang et al., 2023).

This positioning distinguishes the MCP from standard RNN and LSTM practice in two ways. First, mass conservation is explicit at each node and across the network. Second, the internal states and fluxes correspond to physically meaningful quantities such as storage, runoff, evapotranspiration, groundwater exchange, or snow water equivalent. By contrast, the hidden states of conventional RNNs and LSTMs are generally abstract latent features rather than recognized physical stores or fluxes (Wang et al., 2 Oct 2025).

The framework is also intended to span a wide expressivity range. By choice of gate parameterization, the MCP can be as simple as a linear reservoir or as flexible as a universal function approximator. This means that structural simplicity and functional flexibility are separated: one can retain a parsimonious process architecture while allowing the process functions themselves to be learned from data. A plausible implication is that model complexity can be increased in physically meaningful increments instead of by simply adding hidden units.

4. Catchment-scale architectures and process representations

The most developed applications of the MCP are in rainfall-runoff and precipitation-storage-runoff modeling. In catchment-scale hydrological modeling, networks of MCP nodes are used to represent physically meaningful storages such as soil moisture, groundwater, routing stores, and snowpack, with directed links representing quick flow, slow flow, recharge, or bypass. The architecture is designed to be both interpretable and parsimonious, with the explicit aim of discovering a minimal representation adequate for a given catchment (Wang et al., 2024).

A systematic architecture family was examined through six model structures, denoted MA1–MA6. These range from a single storage with a single flow path (MA1) to a three-storage, three-parallel-flow-path configuration (MA6). Intermediate structures include a single storage with two parallel flow paths (MA2), two storages in series (MA3), two storages with two parallel paths (MA4), and a HyMod-like architecture with three storages and two flow paths (MA5). For the Leaf River Basin, the minimal interpretable representation identified from this sequence was a HyMod-like three-state, two-path architecture, and its performance was further improved by adding an input-bypass mechanism and bi-directional groundwater mass exchange (Wang et al., 2024).

Later work developed a process-aware hierarchy within a single MCP storage unit. In that sequence, progressively richer models M1–M5 add bounded soil storage, state-dependent conductivity, variable porosity, infiltration capacity, surface ponding, vertical drainage, and nonlinear water-table dynamics. These variants preserve the same mass-conserving backbone while embedding increasingly explicit hydrological process constraints. The purpose is not simply to increase parameterization, but to test how progressively embedding physically meaningful representations changes predictive skill and interpretability across hydroclimates (Farmani et al., 26 Mar 2026).

5. Empirical performance, diagnostics, and hypothesis testing

The initial proof-of-concept application studied rainfall-runoff dynamics in the Leaf River Basin using a 40-year daily time series with precipitation and PET as inputs and streamflow as output. A single-node MCP with one cell state for soil moisture storage and learned context-dependent gates achieved Median annual KGEss $U_t$4 with worst year KGEss $U_t$5. The learned gating behavior was physically plausible: runoff occurred only when storage exceeded a clear threshold, and conductivity plateaued for large storage, consistent with hydraulic theory. The same study reported that context-dependent output and loss gates substantially improved performance over constant-gate linear reservoirs, particularly during dry years, and that a single-node MCP outperformed ARX, ANN, and RNN benchmarks while matching or exceeding LSTM models with more hidden units (Wang et al., 2023).

The MCP was also used directly for scientific hypothesis testing. Adding a mass-relaxation gate improved baseflow representation in dry years by allowing exchange with the environment, and input bias correction gates learned possible precipitation measurement biases such as rain gauge undercatch at high intensities. Because each added mechanism had a direct hydrological interpretation, model modification was framed as a test of system structure rather than as a purely predictive adjustment (Wang et al., 2023).

Subsequent multi-catchment studies emphasized that performance gains are strongly regime dependent. Across 15 catchments spanning five hydroclimatic regions of the continental United States, progressive process augmentation generally improved skill, with median KGE increasing from ~0.62 (M1) to ~0.77 (M5) and the best MCP configurations approaching a Long Short-Term Memory benchmark at ~0.80. The same study found that MCP outperformed LSTM in 7/15 catchments, that vertical drainage substantially improved skill in arid and snow-dominated basins but reduced performance in rainfall-dominated regions, and that surface ponding had comparatively small effects (Farmani et al., 26 Mar 2026).

A larger-sample extension used 513 US/CONUS catchments and compared single-state soil and snow MCPs with two-state HYDRO-MCP configurations. Model selection favored parsimony using Akaike Information Criterion (AIC). In that study, SOIL-MCP was selected at ~64% of sites, SNOW-MCP at 14%, and HYDRO-MCP at 22%. The resulting OPTMCP achieved median KGE = 0.85, close to a distributed MCP neural network and only modestly below a 5-cell LSTM with KGE = 0.89, while matching or outperforming more complex neural nets at 240/513 catchments when penalizing for complexity (Wang et al., 2 Oct 2025).

6. Broader scope, regionalization, and nomenclature

The MCP has been extended conceptually beyond single-basin rainfall-runoff modeling. One direction is coupling multiple MCP units into larger directed graphs representing snow, soil, groundwater, and routing components in series or in parallel. In the CONUS-scale precipitation-storage-runoff setting, this led to SNOW-MCP, SOIL-MCP, and coupled HYDRO-MCP formulations, including series and bypass configurations. Parallel snow bypass routes often performed better and trained more stably, and the overall program was framed as a step toward regionally adaptive and physically interpretable “models of everywhere” (Wang et al., 2 Oct 2025).

A further extension discussed in the foundational MCP paper is the generalization from mass conservation to coupled mass-energy-information flows. That discussion proposes energy-conserving perceptrons (ECP) and information-flow perceptrons (IFP), as well as mixed MCP/ECP/IFP networks for representing coupled environmental systems. This suggests a broader design principle in which conservation-law-aware computational units serve as the building blocks of differentiable process models (Wang et al., 2023).

A point of nomenclature requires care. In another area of machine learning, MCP denotes the Maximum Cosine Perceptron, a linear classifier derived from the Maximum Cosine Framework. That algorithm optimizes a cosine-based bound on alignment between the current hypothesis and an unknown target hyperplane, shares the Perceptron mistake bound $U_t$6, and was reported to outperform PA and Aggressive ROMMA under single-pass learning on MNIST (Bshouty et al., 2017). Despite the acronym overlap, it is distinct from the Mass-Conserving Perceptron, which is a recurrent, conservation-enforcing unit for geoscientific and hydrological modeling rather than an online linear classifier.

The central idea running through the hydrological MCP literature is therefore not simply that physics can regularize ML, but that conservation laws, process pathways, and interpretable state variables can be encoded directly into the computational unit itself. This suggests a model class in which predictive performance, parsimony, and process diagnosis are treated as joint design objectives rather than as separate stages of analysis.