Quantifying Structural Uncertainty in Physical Models

49 concepts46 papers1 proposal

When we build computational models to simulate physical systems—whether it is a wet friction clutch, a biological regulatory network, or global hydrological cycles—we face two distinct types of ignorance. The first is parametric uncertainty: we know the correct differential equations governing the system, but we do not know the exact physical constants. The second, far more insidious type is structural model uncertainty (or model-form error). This occurs when the mathematical architecture of the model itself is wrong, incomplete, or oversimplified.

As Sankaran Mahadevan notes in ERROR AND UNCERTAINTY QUANTIFICATION AND SENSITIVITY ANALYSIS IN MECHANICS COMPUTATIONAL MODELS, computational models frequently rely on simplified assumptions—such as treating a non-linear material as linear—creating structural discrepancies that escape standard validation. If the equations themselves are an unfaithful representation of reality, no amount of data or parameter tuning can fully resolve the bias.

The Danger of Over-Tuning

Traditional uncertainty quantification (UQ) frameworks focus heavily on data noise and parameter calibration. As highlighted by Apostolos F. Psaros et al. in Uncertainty Quantification in Scientific Machine Learning: Methods, Metrics, and Comparisons, standard approaches struggle to account for structural errors when the underlying physical equations are fundamentally incorrect.

When confronted with a structural mismatch, a common temptation is to force the model to fit the data by aggressively calibrating its parameters. However, Wannes De Groote et al. demonstrate in Inverse parametric uncertainty identification using polynomial chaos and high-order moment matching benchmarked on a wet friction clutch that inverse identification methods often fail when they assume the underlying physical laws are perfectly accurate.

Worse yet, forcing parameters to absorb structural errors can backfire. Andrew White et al. uncover a counter-intuitive trap in The Limitations of Model-Based Experimental Design and Parameter Estimation in Sloppy Systems: forcing poorly constrained parameters to become identifiable can actually destroy a model's predictive power. By targeting the most sensitive regions of an incorrect model, optimization amplifies the impact of structural inaccuracies, turning a useful approximation into an invalid representation.

┌─────────────────────────────────────────────────────────┐
│               True Physical System                      │
└────────────────────────────┬────────────────────────────┘
                             │ Omitted Physics /
                             │ Simplified Geometries
                             ▼
┌─────────────────────────────────────────────────────────┐
│               Flawed Model Structure                    │
└────────────────────────────┬────────────────────────────┘
                             │
                             │ Parameter Tuning Alone
                             ▼
┌─────────────────────────────────────────────────────────┐
│  Overfitted Parameters & Biased/Unreliable Predictions  │
└─────────────────────────────────────────────────────────┘

How Structural Uncertainty Breaks Decisions

If our experimental design or optimization strategies operate under a "perfect-model" assumption, they become fragile and overconfident. In data acquisition, Michael Jordan et al. point out in Robust design of biological experiments that robust experimental design frameworks typically assume the mathematical structure (such as specific chemical reaction formulas) is correct. This leaves data collection strategies highly vulnerable to missing components or incorrect interactions in the model diagram.

Similarly, in Bayesian Optimization (BO), surrogate models like Gaussian Processes can become dangerously overconfident if they ignore structural imprecision (Accounting for Gaussian Process Imprecision in Bayesian Optimization). As Zishi Zhang et al. show in Exploring and Exploiting Model Uncertainty in Bayesian Optimization, if the chosen model class cannot represent the true functional form, the optimization process will likely converge to sub-optimal results regardless of how much data is collected.

Strategies for Structural Inference and Correction

To resolve this, researchers are shifting from tuning parameters to learning, marginalizing, and correcting the model structure itself.

Local Dynamic Operators

Instead of searching an abstract, infinite space of possible equations, Anthony M. DeGennaro et al. propose a clever reparametrization in MODEL STRUCTURAL INFERENCE USING LOCAL DYNAMIC OPERATORS. By representing the local structure of partial differential equations as a "Local Dynamic Operator" (LDO), they transform the abstract problem of structural inference into a manageable parametric search space, allowing physics models to be calibrated directly to empirical data.

Hybrid Physics-ML Frameworks

Another powerful approach is the "gray-box" model, which embeds machine learning components to capture omitted physical behaviors while retaining known physical laws. For example, Shailesh Garg et al. introduce a hybrid framework in Physics-integrated hybrid framework for model form error identification in nonlinear dynamical systems that uses Bayesian filters to isolate hidden forces and Gaussian Process regression to learn their structural patterns.

Where the physics is entirely missing or highly uncertain, frameworks like the one developed by Kevin Course et al. in State estimation of a physical system with unknown governing equations bypass rigid differential equation solvers entirely, tracking the system's state while simultaneously learning its underlying governing equations from noisy, indirect data.

Structural Marginalization and Ensembles

Rather than committing to a single imperfect model structure, we can treat the structure itself as a random variable or combine multiple viewpoints:

By acknowledging that our mathematical formulas are approximations, these methods allow scientific machine learning to move past the rigid boundaries of classical equations, ensuring that our error bars reflect not just what we don't know about our parameters, but what we don't know about our models.

Key Foundational Work

Robust design of biological experiments (2005) — contributed 2 concept elements to this direction.
Model Identification and Assessment Based on Model Predicted Output (1998) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_structural_model_uncertainty_ad8c0e66,
  title = {structural model uncertainty},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/structural-model-uncertainty/rdr_ad8c0e666fa443be8f52e82deddb2835},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_structural_model_uncertainty_ad8c0e66_v1,
  title = {structural model uncertainty},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/structural-model-uncertainty/rdr_ad8c0e666fa443be8f52e82deddb2835/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake