Hybrid Uncertainty Modeling in Physics and Data

Updated 6 February 2026

Physics-based, data-driven, and hybrid uncertainty modeling are integrated strategies that quantify and propagate uncertainty in complex systems.
Physics-based methods use governing equations and spectral techniques, while data-driven approaches employ machine learning to capture empirical trends.
Hybrid frameworks combine physical models with data corrections, enhancing prediction accuracy and efficiency in scientific simulations.

Physics-based, data-driven, and hybrid uncertainty modeling encompasses a spectrum of strategies for quantifying, propagating, and mitigating uncertainty in complex scientific and engineering systems. These approaches are distinguished by their treatment of underlying governing principles, the integration of empirical observation, and their ability to generate physically consistent, computationally efficient, and robust predictions with rigorously quantified confidence. This article details the foundational frameworks, methodological implementations, and comparative performance of physics-based, data-driven, and hybrid models for uncertainty quantification (UQ), emphasizing recent advances and unified architectures across representative domains.

1. Foundational Principles and Model Classes

Physics-based uncertainty modeling relies on the explicit propagation of uncertainty through governing equations, often involving partial differential equations (PDEs) or system-specific operators derived from first principles. The primary sources of uncertainty addressed include parameter variability, stochastic boundary conditions, and model-form errors introduced by physical simplification or closure approximations. Methodologies such as generalized polynomial chaos (gPC), spectral or stochastic Galerkin projection, and intrusive or non-intrusive collocation are widely utilized for rigorous propagation and error control at the PDE level (Mittal et al., 2014).

Data-driven uncertainty modeling, in contrast, abstracts the underlying process through statistical or ML surrogates, capturing correlations, trends, and responses present in empirical observations. Uncertainty quantification in this paradigm generally entails distributional regression (e.g., quantile regression, deep ensembles, Bayesian neural networks, Gaussian process regression), with uncertainty reflecting both epistemic (data limitation, model-form) and aleatoric (input/observation noise) sources. The resulting models offer high flexibility but may suffer from poor extrapolation, unphysical predictions, and opaque uncertainty attributions (Ratn et al., 11 Dec 2025, Furlong et al., 26 Feb 2025).

Hybrid approaches seek to integrate physical knowledge with data-driven inference, leveraging the structure, consistency, and inductive bias provided by physics while utilizing data-driven corrections or augmentations to address unmodeled, complex, or uncertain phenomena. These frameworks span residual learning (correction of physics solvers with supervised surrogates), physics-informed deep kernel Gaussian processes, modularized UQ with module-level method selection, and adaptive mixture strategies that switch between physics and data-driven solvers based on local uncertainty estimates (Mittal et al., 2014, Storm et al., 23 Apr 2025, Ratn et al., 11 Dec 2025).

2. Module-Based and Global Frameworks for Uncertainty Quantification

Module-based hybrid frameworks, typified by (Mittal et al., 2014), enable each physics module in a multi-physics or multi-scale simulation to implement the UQ method most suited to its mathematical structure and computational constraints: intrusive stochastic Galerkin for low-dimensional, highly resolved blocks; non-intrusive regression or spectral projection for legacy or black-box codes; and semi-intrusive approaches for reduced-sampling via solver augmentation. To facilitate coupling, restriction and prolongation operators map between global and module-local generalized polynomial chaos (gPC) coefficients, supporting block Gauss–Seidel iteration until convergence in the joint stochastic-physical state.

A representative summary of comparison across approaches:

Model Class	Main UQ Propagation	Scalability	Intrusiveness	Interpretability
Physics-based	gPC/spectral, PDE	Limited (curse of dim.)	High	Explicit, physically grounded
Data-driven	ML regression, GP	High	None	Often opaque, may not be physical
Hybrid (module-based)	Modular (mixing), gPC+ML	Moderate-High	Selective	Adaptable, interpretable at module level

Case studies such as thermally-driven cavity flow demonstrate that the hybrid modular strategy (e.g., intrusive flow + non-intrusive heat) achieves equivalent accuracy and reduced computational time (2–3× speedup) relative to monolithic intrusive solvers, and enables legacy code reuse while isolating high-dimensional stochastic expansions to those modules for which they are essential (Mittal et al., 2014).

3. Hybrid Residual Correction and Data-Efficient Uncertainty Decomposition

Residual learning is a dominant paradigm in hybrid models. A physics-informed model acts as a primary predictor, and a secondary data-driven model is trained on the residuals between observed quantities and physical model outputs. Gaussian process regression (GPR) and neural networks are the canonical choices for this secondary surrogate (Ratn et al., 11 Dec 2025, Furlong et al., 26 Feb 2025, Cillis et al., 30 Jan 2026).

This pipeline enables decomposable uncertainty quantification:

Epistemic uncertainty: Estimated from the GPR posterior variance, deep ensemble spread, or Bayesian weight covariance, reflecting model-form and data-scarcity uncertainties.
Aleatoric uncertainty: Quantified via analytic propagation (e.g., Delta method for multivariate input covariance—linearization of model outputs with respect to noisy or correlated inputs (Ratn et al., 11 Dec 2025), or conformal quantile regression (Gijón et al., 11 Feb 2025, Furlong et al., 26 Feb 2025)) and added to the epistemic term for full predictive intervals.

Example: In forward osmosis modeling, the residual-GPR hybrid achieves MAPE=0.26%, R²=0.999 with total variance $\sigma_{\mathrm{total}}^2(z_*) = \sigma_{\mathrm{epistemic}}^2(z_*) + \sigma_{\mathrm{aleatoric}}^2(z_*)$ ; finite-difference or autodiff provides the necessary Jacobians for analytic input variance propagation (Ratn et al., 11 Dec 2025). Similarly, in wind turbine power, explainable uncertainty (via conformal quantile intervals) is complemented by SHAP value analysis for interpretability (Gijón et al., 11 Feb 2025).

4. Uncertainty-Driven Model Selection, Mixture, and Adaptivity

Advanced hybrid frameworks dynamically select between, or blend, tailored surrogate and high-fidelity models by monitoring the estimated local uncertainty. The uncertainty-driven phase field approach, as in (Storm et al., 23 Apr 2025), introduces a scalar field $\phi(x)\in [0,1]$ that interpolates the constitutive response between a fast GP surrogate and a high-fidelity solver. The phase evolution is governed by a Landau-esque free-energy functional with uncertainty-driven terms, ensuring both convergence (diffuse interface) and computational efficiency by minimizing high-fidelity calls in well-extrapolated regions.

Table: Mixture-model strategies

Mixing Scheme	Criteria	Transition Mechanism	Key Parameters
Uncertainty-driven PF	Surrogate UQ $U(x)$	Phase-field PDE/Δ $t$	Interface width $\epsilon$ , threshold $b$
Local switching	Pointwise UQ	Thresholding/logic	Cut-off $\tau$ , no smoothing

Empirical studies show that, for the same MAE, phase field mixture often requires only 35–50% the number of high-fidelity evaluations compared to brute-force HF everywhere, with smooth transitions and mechanical stability even under mesh or time-step changes (Storm et al., 23 Apr 2025).

5. Physics-Constrained Machine Learning and Deep Kernel Gaussian Processes

Physics-constrained Gaussian process regression, especially with deep kernels, provides a systematic mechanism to incorporate PDE constraints and physical laws into the data-centric UQ framework (Chang et al., 2022). The loss function combines GP marginal likelihood with a weighted physics residual term (e.g., Boltzmann–Gibbs penalty for the governing PDEs), enforced over labeled and unlabeled data points. The covariance kernel itself is parameterized by a neural network encoder to capture high-dimensional correlations efficiently.

Posterior predictions inherit the tractable UQ of GP surrogates, while the physics penalty ensures consistency (often reducible to orders of magnitude less labeled data than unconstrained DNN surrogates). This approach has been demonstrated to reach target accuracy and reliable UQ for surrogate modeling of 2D high-dimensional random PDEs with O(10³) training samples (Chang et al., 2022).

6. Operator Inference and ROM-Based Hybrid UQ in Multi-Physics

Operator-inference-based reduced-order models (ROMs) provide a powerful strategy for efficient UQ in multi-physics and multi-scale settings with high-dimensional state spaces and parametric uncertainty (Behnoudfar et al., 4 Oct 2025). Here, low-dimensional subspaces (POD modes) are learned from high-fidelity simulation snapshots, and projected dynamical operators (affine with respect to targeted uncertain parameters) are inferred via regression. Once constructed, these ROMs enable rapid uncertainty propagation (Monte Carlo over parametric distributions) orders of magnitude faster than the underlying high-fidelity code, with demonstrated accuracy losses below 1–2% for test quantities of interest.

Applications to reacting flows in porous media, ablative heat shields, and solid/gas-phase combustion validate the approach: with $360\times$ – $380\times$ speedup, full distributional UQ is enabled within practical computational budgets for design and probabilistic risk analysis (Behnoudfar et al., 4 Oct 2025).

7. Comparative Performance, Scalability, and Generalization

Empirical and case-study-driven analysis demonstrates:

Physics-based UQ methods achieve best accuracy for smooth, well-posed systems but suffer from scalability limits and code-intrusiveness in complex, high-dimensional problems (Mittal et al., 2014, Krannichfeldt et al., 23 Jul 2025).
Data-driven UQ is flexible, scalable, and often achieves superior fit in the presence of rich datasets, but extrapolation and physical consistency remain problematic; uncertainties may be significantly over- or under-estimated out-of-sample (Gijón et al., 11 Feb 2025, Xiao et al., 2015).
Hybrid methods, particularly residual correction and modular or phase-field mixture architectures, exploit the structure of the physics to minimize data requirements, maintain interpretability, and provide robust and calibrated predictive intervals under both epistemic and aleatoric uncertainty (Ratn et al., 11 Dec 2025, Storm et al., 23 Apr 2025, Gijón et al., 11 Feb 2025, Krannichfeldt et al., 23 Jul 2025). In digital twin, engineered, or out-of-distribution scenarios, hybrid surrogates (e.g., residual-QNN in building energy, PGML in aerodynamic forces) consistently outperform both pure-physics and pure-ML models in terms of coverage, sharpness, and calibration (Pawar et al., 2021, Krannichfeldt et al., 23 Jul 2025).

Hybrid frameworks further admit extensibility to transfer learning, multi-fidelity coupling, and adaptive sampling for active learning, suggesting a broad and robust foundation for future uncertainty-aware scientific machine learning (Cillis et al., 30 Jan 2026, Pawar et al., 2019).

References

"A Flexible Uncertainty Quantification Framework for General Multi-Physics Systems" (Mittal et al., 2014)
"Hybrid Physics-ML Model for Forward Osmosis Flux with Complete Uncertainty Quantification" (Ratn et al., 11 Dec 2025)
"Integrating Physics and Data-Driven Approaches: An Explainable and Uncertainty-Aware Hybrid Model for Wind Turbine Power Prediction" (Gijón et al., 11 Feb 2025)
"Mixing Data-Driven and Physics-Based Constitutive Models using Uncertainty-Driven Phase Fields" (Storm et al., 23 Apr 2025)
"Hybrid Physics-Data Enrichments to Represent Uncertainty in Reduced Gas-Surface Chemistry Models for Hypersonic Flight" (Bandy et al., 9 Sep 2025)
"Integrating Physics-Based and Data-Driven Approaches for Probabilistic Building Energy Modeling" (Krannichfeldt et al., 23 Jul 2025)
"Uncertainty quantification of reacting fluids interacting with porous media using a hybrid physics-based and data-driven approach" (Behnoudfar et al., 4 Oct 2025)
"Quantifying and Reducing Model-Form Uncertainties in Reynolds-Averaged Navier-Stokes Simulations: A Data-Driven, Physics-Based Bayesian Approach" (Xiao et al., 2015)
"Physics-Based Hybrid Machine Learning for Critical Heat Flux Prediction with Uncertainty Quantification" (Furlong et al., 26 Feb 2025)
"A hybrid data driven-physics constrained Gaussian process regression framework with deep kernel for uncertainty quantification" (Chang et al., 2022)
"Hybrid analysis and modeling for next generation of digital twins" (Pawar et al., 2021)
"Data-driven recovery of hidden physics in reduced order modeling of fluid flows" (Pawar et al., 2019)
"DiffHybrid-UQ: Uncertainty Quantification for Differentiable Hybrid Neural Modeling" (Akhare et al., 2023)
"Exploring Efficient Quantification of Modeling Uncertainties with Differentiable Physics-Informed Machine Learning Architectures" (Oddiraju et al., 23 Jun 2025)
"Hybrid physics-data-driven modeling for sea ice thermodynamics and transfer learning" (Cillis et al., 30 Jan 2026)