Scientific machine learning in Hydrology: a unified perspective

This review presents the first unified conceptual framework for scientific machine learning in hydrology, organizing fragmented methodological families into a coherent structure to clarify novelty, foster cumulative progress, and guide future research directions.

The Gist

Imagine you are trying to predict the weather, the flow of a river, or how much water is in a lake. For decades, scientists have used two main tools to do this:

1. The "Physics Book" Approach: Using complex math equations based on the laws of nature (like gravity and conservation of mass). It's accurate but slow, expensive to run, and sometimes misses the messy details of the real world.
2. The "Data Detective" Approach: Using Artificial Intelligence (Machine Learning) to look at past data and guess the future. It's fast and flexible but can be "hallucinated" (make up nonsense) if it hasn't seen enough data, and it doesn't understand why things happen.

Scientific Machine Learning (SciML) is the new super-tool that tries to marry these two. It's like giving the Data Detective a copy of the Physics Book so they can learn faster and make smarter guesses.

This paper by Adoubi Vincent De Paul Adombi is a map and a guidebook for this new super-tool. The author noticed that scientists are building these hybrid tools in many different, confusing ways, often reinventing the wheel. He organizes them into four distinct families, explaining how they work, where they fail, and where they can go next.

Here is a simple breakdown of the four families using everyday analogies:

---

1. The "Strict Coach" (Unified Physics-Informed Machine Learning)

The Analogy: Imagine a student taking a test. Usually, they just try to get the right answers based on what they memorized. But in this method, the teacher (the Physics Law) sits next to them and whispers, "Hey, you can't say the water flows uphill!" or "Remember, water can't just disappear!"

• How it works: The AI is forced to obey the laws of physics while it is learning. If the AI makes a prediction that breaks the laws of nature, it gets a "penalty" (a bad grade) and has to try again.
• The Good: It's very reliable even when data is scarce because the physics rules act as a safety net.
• The Bad: It's computationally heavy (like running a marathon while carrying a heavy backpack). Also, once the test is over, the "whispering teacher" leaves. If the river changes course tomorrow, the AI might not know how to adapt without retraining.

2. The "Mentor with a Cheat Sheet" (Unified Physics-Guided Machine Learning)

The Analogy: Imagine you are hiring a brilliant but inexperienced intern (the AI) to predict river flow. You give them a "Cheat Sheet" generated by a physics expert. The intern doesn't have to be the expert, but they can look at the Cheat Sheet to help them make better guesses.

• How it works: Instead of forcing the AI to obey physics rules, you feed the AI the results of a physics simulation as extra information. The AI learns to combine its own data patterns with the expert's simulation.
• The Good: It's very flexible and can learn complex patterns that pure physics might miss.
• The Bad: If the "Cheat Sheet" (the physics simulation) is wrong or biased, the AI will learn those mistakes too. It's only as good as the expert who wrote the cheat sheet.

3. The "Specialized Team" (Hybrid Physics-Machine Learning)

The Analogy: Imagine a construction crew building a house. You have a master architect (Physics) who knows the structural rules, and a creative interior designer (AI). Instead of forcing the designer to follow every structural rule, you let them work together:

• Additive: The architect builds the frame, and the designer fixes the cracks.

• Embedded: The designer builds a room inside the architect's blueprint.

• Replacement: You replace a slow, clunky part of the architect's plan with a fast, smart AI module.

• How it works: This approach keeps the physics model and the AI model separate but connected. They talk to each other to produce the final result.

• The Good: It's great for fixing specific problems in old models without rebuilding everything from scratch.

• The Bad: It can be tricky to get them to talk to each other correctly. Sometimes the AI just "patches" the physics model to look good on paper without actually fixing the underlying physics problem.

4. The "Archaeologist" (Physics Discovery)

The Analogy: Imagine finding a pile of ancient, broken pottery shards (data) but no instructions on how they fit together. The Archaeologist (AI) tries to figure out the original shape of the pot and even write a new rulebook on how the pot was made, just by looking at the shards.

• How it works: Instead of using known physics, the AI looks at the data and tries to invent the mathematical equations that describe the system. It's like the AI writing its own textbook.
• The Good: It can discover new laws of nature or hidden processes that humans haven't thought of yet.
• The Bad: It's risky. The AI might find a pattern that fits the data perfectly but makes no sense in the real world (like a magic spell that works in a video game but not in reality). It needs a lot of clean data to work well.

---

The Big Picture: Why This Matters

The author argues that right now, scientists are all speaking different languages. One person calls their method "Physics-Guided," another calls it "Hybrid," and they don't realize they are doing similar things.

This paper provides a unified dictionary. It says: "Here is how we all fit together."

The Future Outlook:

• The Challenge: These tools are still expensive to run and can be fragile if the data is messy.
• The Opportunity: If we can make these tools faster and more robust, we could predict floods, manage droughts, and understand climate change much better than we do today. We could move from just "guessing" the future to truly "understanding" it.

In short: This paper is the instruction manual for the next generation of water scientists, teaching them how to build AI that respects the laws of nature while learning from the messy reality of the world.

Technical Summary

1. Problem Statement

Hydrological research faces significant challenges due to limited/noisy observational data, the high computational cost of traditional numerical simulators, and the persistent gap between physical interpretability and predictive performance. While Scientific Machine Learning (SciML) has emerged as a paradigm to integrate physical knowledge with data-driven modeling, the field has become fragmented.

• Fragmentation: Multiple methodological families (Physics-Informed, Physics-Guided, Hybrid, and Physics Discovery) have evolved in parallel with heterogeneous approaches.
• Lack of Coordination: Contributions are often presented in isolation without a unified conceptual framework, making it difficult to assess novelty, identify gaps, or lower the barrier to entry for researchers from different disciplines (e.g., hydrologists vs. ML experts).
• Need for Unification: There is an urgent need for a structured overview that categorizes these diverse methods, clarifies their internal logic, and fosters cumulative progress.

2. Methodology: A Unified Framework

The paper proposes a unified methodological framework for four distinct families of SciML approaches in hydrology. Rather than listing every existing method, the author abstracts them into modular architectures and composite loss functions to reveal recurring design patterns.

A. Unified Physics-Informed Machine Learning (UPIML)

• Core Concept: Integrates physical laws (PDEs, boundary conditions) as constraints within the training objective (loss function) of neural networks.
• Architecture: Composed of State Modules (SMs) predicting physical states (e.g., water level) and Parameterization Modules (PMs) estimating latent parameters (e.g., hydraulic conductivity).
• Loss Function: A composite loss ($\mathcal{L}_{total}$) balancing:
  • Data Misfit: Agreement with observations.
  • Physics Residuals: Penalties for violating governing equations (using Automatic Differentiation).
  • Interface/Continuity: Constraints between coupled modules.
  • Expert Knowledge: Empirical rules or inequality constraints.
  • Regularization: Smoothness and weight decay.
• Special Cases: Includes standard PINNs (single state, no parameters), multi-state domain decomposition, and architectures with embedded parameter inference.

B. Unified Physics-Guided Machine Learning (UPGML)

• Core Concept: Uses outputs from deterministic or stochastic physics-based simulators as features or intermediate representations to guide the learning process, rather than enforcing constraints in the loss.
• Architecture: A five-stage pipeline:
  1. Physical Feature Generation (PFgM): Runs a physics simulator to generate signals.
  2. Input Encoding (IEM): Fuses raw forcings with physical signals.
  3. Latent Representation Learning (LRLM): Learns system dynamics (e.g., via LSTM/MLP).
  4. Latent Fusion (LFM): Injects physical features into hidden layers (optional deep integration).
  5. Output Mapping (OMM): Generates final predictions.
• Special Cases: Ranges from shallow fusion (concatenating physics outputs as inputs) to deep latent fusion (injecting physics into hidden layers).

C. Hybrid Physics-Machine Learning

• Core Concept: Maintains a clear separation between physical modules and data-driven components, allowing for modular integration.
• Three Categories:
  1. Additive Learning: ML models predict the residual error between a physics model and observations (Physics + ML Residual).
  2. Physics-Embedded ML: Physics models are embedded as differentiable layers within a neural network, followed by a post-processing ML block.
  3. Submodule Replacement: Specific components of a physics-based model (e.g., infiltration or evapotranspiration) are replaced by trainable ML surrogates while preserving the global model structure.

D. Physics Discovery

• Core Concept: Directly discovers unknown physical laws or model structures from data.
• Three Categories:
  1. Symbolic Regression: Searches for explicit algebraic equations (e.g., AI Feynman) to describe variable relationships.
  2. Stochastic Universal Partial Differential Equations (SUPDE): Models systems using differential equations that include learnable deterministic components and stochastic noise terms.
  3. Conceptual Model Discovery: Automatically learns the structure of "bucket-type" models (reservoirs and fluxes). Examples include DeepDiscover (infers architecture via recurrent units), Mass-Conserving Perceptron (MCP) (enforces mass balance via gates), and Deep Process Learning (DPL) (integrates physical reasoning into RNNs).

3. Key Contributions

1. First Unified Review: This is the first dedicated synthesis of SciML specifically for hydrology, organizing a fragmented literature into four coherent families.
2. Modular Abstraction: The paper defines generic architectures (UPIML, UPGML) that encompass specific existing methods as special cases, providing a common language for the field.
3. Critical Analysis of Limitations: It systematically identifies bottlenecks for each family:
   • UPIML: High computational cost, sensitivity to initialization, and poor generalization to changing boundary conditions (static PDE solver behavior).
   • UPGML: Dependence on the accuracy of the underlying physics simulator; if the physics model is biased, the ML model inherits that bias.
   • Hybrid: Issues with non-differentiable legacy codes and inference dependency (ML needing physics outputs at runtime).
   • Discovery: Sensitivity to noise, combinatorial explosion in search space, and issues with model identifiability (non-uniqueness of solutions).
4. Future Roadmap: Proposes specific directions for advancement, such as differentiable surrogates, adaptive sampling, causal weighting in loss functions, and automated diagnostic tools for model coupling.

4. Results and Evidence

The paper does not present new experimental results but synthesizes findings from representative case studies to validate the frameworks:

• UPIML: Cites studies (e.g., Ali et al., 2024; Luo et al., 2024) showing successful groundwater and river network modeling but highlighting computational overhead compared to traditional solvers.
• UPGML: References work (e.g., Young et al., 2017; Houénafa et al., 2025) demonstrating improved streamflow prediction in data-scarce regions by fusing HEC-HMS or stochastic rainfall-runoff outputs with ML.
• Hybrid: Highlights studies (e.g., Cho & Kim, 2022; Adombi et al., 2024) where residual learning or submodule replacement improved NSE/KGE scores over pure physics or pure ML models.
• Discovery: Discusses DeepDiscover (Adombi, 2025) achieving high NSE (0.68) and KGE (0.70) on CAMELS-US basins, outperforming traditional conceptual models and pure CNNs, while maintaining physical interpretability.

5. Significance

• Conceptual Clarity: By unifying disparate methods, the paper lowers the entry barrier for hydrologists and ML researchers, facilitating interdisciplinary exchange.
• Cumulative Progress: The framework allows researchers to identify where meaningful advances can still be made (e.g., moving from static UPIML to dynamic, transferable architectures).
• Operational Relevance: It highlights the transition from purely theoretical SciML to practical applications, emphasizing the need for scalability, robustness to data scarcity, and the ability to handle changing climate conditions.
• Future Guidance: The review serves as a roadmap for developing next-generation hydrological models that are not only accurate but also physically consistent, interpretable, and computationally efficient.