Physics-Guided Machine Learning

• PGML is a field that integrates explicit physical constraints into machine learning models to enforce conservation laws and reduce uncertainty.
• It employs hybrid methodologies such as hard-embedded constraints and physics-based regularization to enhance model robustness and interpretability.
• PGML has been applied successfully in hydrology, fluid mechanics, photonics, and quantum simulation to accelerate and improve predictive modeling.

Physics-Guided Machine Learning (PGML) integrates domain-specific physical laws, structures, or simplified process models into ML frameworks to improve generalization, reduce uncertainty, and ensure physical consistency, particularly in scientific and engineering applications. By leveraging physics embeddings in ML architectures, loss functions, input representations, or data workflows, PGML aims to overcome the limitations of both purely data-driven models and rigid physics-based approaches. PGML has been successfully implemented across a variety of domains, including hydrology, fluid mechanics, photonics, environmental science, turbulence modeling, and quantum simulation.

1. Conceptual Foundations and Theoretical Principles

Physics-Guided Machine Learning can be formally understood as a subclass of Physics-Informed Machine Learning (PIML), emphasizing structural or architectural imposition of physical knowledge rather than relying solely on soft-penalty terms in the loss function. The foundational premise is that learning algorithms can be regularized or augmented using domain-specific constraints such as conservation laws (mass, energy, momentum), analytic solutions from simplified models, or invariances (e.g., symmetry, stability) (Nghiem et al., 2023).

A general PGML model includes a data-fidelity term and a physics-consistency term in the training objective: $$L(\theta) = L_\text{data}(\theta) + \lambda\,L_\text{physics}(\theta)$$ where $L_\text{data}$ measures prediction error and $L_\text{physics}$ penalizes violations of physical constraints, with $\lambda$ controlling the trade-off.

These constraints can be encoded as:

• Hard embedding: Imposing structural design, e.g., mass/energy balance equations as layers or analytic head functions.
• Soft penalties: Physics residuals included in the loss function (e.g., PDE residuals, conservation-law mismatches).
• Feature augmentation: Injection of features derived from low-fidelity or reduced-(order) physical models at intermediate or input layers of deep networks (Pawar et al., 2021, Pawar et al., 2020, Pawar et al., 2021).

2. Key PGML Methodologies

Several methodological patterns typify PGML implementations:

2.1 Structural Hybridization

Hybrid models architecturally integrate physics-based modules (e.g., conceptual or empirical process models) with ML surrogates. For instance, in hydrological modeling, the abcd conceptual model provides state-update equations and process decompositions (e.g., mass-balance, actual evapotranspiration, runoff), with ML surrogates predicting intermediate components while embedded in a hard-coded mass-balancing architecture (Esmaeilzadeh et al., 2024). Generalizations include:

• Neural ODEs / UDEs: Split ODE right-hand-sides into physics-based and neural components.
• Physics-based heads: Analytical, parameterized physical models (e.g., NRTL for activity coefficients) serve as output layers (Winter et al., 2022).

2.2 Physics-Based Regularization

PGML frequently penalizes violation of known physical laws in the loss function. Example patterns include:

• Mass/energy conservation: Compare network predictions with analytic conservation updates, penalizing residuals (Esmaeilzadeh et al., 2024, Prasad et al., 2024, Yu et al., 10 Feb 2025).
• PDE residuals: For PINNs or surrogate PDE solvers, include squared norms of governing-equation residuals at collocation points (Nghiem et al., 2023, Lynch et al., 24 Feb 2025).
• Thermodynamic consistency: Penalize violations of the Gibbs–Duhem relation in activity-coefficient prediction (Winter et al., 2022).

2.3 Feature and Latent Variable Injection

Physically meaningful quantities—derived from analytical or simplified models—are injected as inputs or concatenated into the latent space of a network. This approach constrains the model’s hypothesis space and improves epistemic robustness (Pawar et al., 2020, Pawar et al., 2021, Pawar et al., 2021). For example, aerodynamic prediction networks inject Hess–Smith panel-method outputs at an intermediate hidden layer, reducing prediction variance by 75% (Pawar et al., 2021).

2.4 Physics-Guided Data Workflows

PGML data pipelines emphasize strategic selection, augmentation, and correction of training data informed by physics. Key practices include:

• Active sample reweighting: Assign highest weights to samples with high physical consistency and high model uncertainty to enhance robustness against noise and physical inconsistency in data-sparse, real-world applications (Jiang et al., 2024).
• Foundation model pretraining: Massively synthetic, physics-consistent datasets enable pre-training of backbone architectures, which are fine-tuned via real-world data under multi-task physics-consistent objectives (Yu et al., 10 Feb 2025).

3. Applications and Domain-Specific Implementations

PGML frameworks have delivered significant improvements across multiple application areas:

3.1 Environmental and Hydrological Modeling

PGML models in hydrology and climate applications enforce process connectivity and conservation via embedded surrogates for intermediate states (soil moisture, snowpack, ET) and by integrating known mass-balance structures (abcd or SWAT). In streamflow forecasting, PGML achieved Nash-Sutcliffe efficiency (NSE) gains of 0.03–0.08 over standalone ML, and reductions in RMSE for both Q and ET (Esmaeilzadeh et al., 2024, Khandelwal et al., 2020). In snow-ice emulation, physics-guided LSTM delivered nearly 10,000× speedup while reducing RMSE and maintaining mass consistency (Prasad et al., 2024).

3.2 Fluid Mechanics and Surrogate Modeling

PGML surrogates for fluid dynamics, turbulence models, and reduced order flow solvers inject Galerkin-reduced system predictions to improve generalization, enable extrapolation, and sharply reduce uncertainty in closure modeling. For vorticity transport, PGML variational multiscale closure models reduced modal coefficient MSE by more than an order of magnitude over classical ROMs (Ahmed et al., 2022). Hybrid ML/Eigenspace Perturbation in turbulence reduces TKE error by 1–3 orders of magnitude, preserving physical realizability (Chu et al., 7 Nov 2025).

3.3 Photonics and Electromagnetics

Hierarchical convolutional architectures combined with embedded Maxwell residual constraints enable surrogate models for photonic field distributions that require an order of magnitude less labeled data, leapfrogging black-box networks in both accuracy and extrapolation to new doping levels (Lynch et al., 24 Feb 2025). Effective medium behaviors emerge automatically as latent representations in core CNN modules.

3.4 Process Systems and Thermodynamics

In chemical thermodynamics, Transformer-based PGML models map molecular input representations to thermodynamically consistent outputs by embedding parameterized NRTL equations as NN heads and explicitly enforcing the Gibbs–Duhem relation during training (Winter et al., 2022).

3.5 Energy Yield and Planet-Scale Science

PGML approaches that combine physically-clustered climate zones (PVZones), simulator-aligned input variable selection, and scalable data processing yield sub-5% error in global photovoltaic output forecasting, replacing high-cost simulation with millisecond inference (Jahangir et al., 2024).

3.6 Quantum Simulation

Physics-guided generative models, such as PIGen-SQD, filter and anchor quantum hardware-sampled determinants using low-order perturbation theory, then use iterative ML-driven expansion (RBMs) to recover near-exact states at <10% of the computational cost and search space, maintaining chemical accuracy in strongly correlated problems (Patra et al., 7 Dec 2025).

4. Evaluation, Uncertainty Quantification, and Physical Consistency

PGML models are assessed by both standard ML error metrics (RMSE, NSE, MAPE, F1, ROC AUC) and by physics-related scores (mass/energy-balance residuals, adherence to constraints). Robustness and reliability are typically demonstrated by:

• Reduction in prediction uncertainty: E.g., 75% lower ensemble prediction variance in aerodynamic surrogates (Pawar et al., 2021).
• Improved extrapolation/generalization: Near-halving of test error in wind-farm power predictions when physics-based features are incorporated (Zehtabiyan-Rezaie et al., 2022).
• Physical plausibility: PGML surrogates verify mass-balance, thermodynamic, or turbulence realizability constraints by design or via penalty minimization.
• Data efficiency: PGML methods consistently achieve higher fidelity with less training data or simulation expense by leveraging physics-motivated input selection, constraints, or synthetic data expansions (Jahangir et al., 2024, Lynch et al., 24 Feb 2025).

5. Limitations, Open Challenges, and Future Directions

PGML’s limitations and research frontiers include:

• Domain specific tuning: Architectures and regularization coefficients (e.g., $\lambda$ in loss functions) often require bespoke selection or cross-validation; automation of this is an open challenge (Esmaeilzadeh et al., 2024).
• Generalization scope: Demonstrated gains often concern spatial or scenario generalization, but robustness under multi-scale or non-stationary conditions (e.g., distributed urban drainage, three-dimensional turbulence) demands further study (Chu et al., 7 Nov 2025, Palmitessa et al., 2022).
• Physical constraint scope: Many current PGML models apply soft penalties; extension to hard physics enforcement or full nonlinear operator embeddings remains limited (Prasad et al., 2024).
• Hybrid system integration: Scalability to coupled, multi-physics or multi-modal scientific systems (e.g., aquatic–atmospheric–terrestrial) is an active research direction, as is extending foundation-model pretraining to more complex simulators (Yu et al., 10 Feb 2025).
• Uncertainty quantification: Integrating rigorous epistemic and aleatoric uncertainty quantification with physics-guided architectures is an area of ongoing development (Sedehi et al., 2023).
• Transferability and meta-learning: Generalizing pretrained PGML backbones or weights across scientific domains or system classes is in early stages.

6. Representative Example: Table of Core PGML Components

Application Domain	PGML Embedding Type	Physical Constraint / Model
Hydrology	Two-stage surrogate, hard constraints	abcd conceptual mass-balance equations (Esmaeilzadeh et al., 2024)
Photonics	Hierarchical CNN, loss penalties	Maxwell residual constraints (Lynch et al., 24 Feb 2025)
Wind Farms	Feature injection at input	Park / Gaussian wake model efficiency (Zehtabiyan-Rezaie et al., 2022)
Thermodynamics	Model-based analytic NN head	NRTL equations + Gibbs–Duhem (Winter et al., 2022)
Turbulence	ML modulation of operator magnitudes	EPM realizability of Reynolds stresses (Chu et al., 7 Nov 2025)

7. Impact and Broader Significance

Physics-Guided Machine Learning represents a paradigm shift in scientific computing by uniting the interpretability, structure, and generality of first-principles models with the flexibility and expressivity of modern machine learning. PGML advances the data efficiency, extrapolative reliability, and physical consistency of AI-assisted modeling in domains where traditional ML underperforms due to limited real data or the need for rigorous scientific constraint satisfaction. Ongoing success across simulation acceleration, digital twins, environmental forecasting, turbulence, materials design, and quantum computing illustrates the potential and extensibility of PGML methodologies, with open challenges defining the next decade’s research agenda (Nghiem et al., 2023, Yu et al., 10 Feb 2025, Pawar et al., 2021).