Resolving Unmeasurable Parameters in Physical Mechanism Models

22 concepts14 papers1 proposal

Physics-based mechanism models—such as thermodynamic equations, thermal networks, and diffusion laws—are exceptionally powerful tools. They leverage fundamental conservation laws to predict how a system will behave over time. However, when deployed in real-world industrial and engineering environments, these models frequently hit a hard wall: they rely on internal physical parameters, coefficients, or localized states that are completely unmeasurable.

Whether it is the internal junction temperature of an operating semiconductor, the specific heat capacity of a substance undergoing a chaotic phase change, or the diffusion constants inside a high-temperature furnace, directly probing these variables is often physically impossible, prohibitively expensive, or highly destructive.

To bridge this gap, a powerful research direction has emerged. Instead of relying on direct physical sensing, researchers are combining physics-based constraints with data-driven optimization and machine learning. By treating unmeasurable coefficients as latent variables and running physical equations "backward," these hybrid frameworks infer hidden internal states from easily accessible external observations.

The Measurement Gap in Pure Physics Models

To understand why this hybrid approach is necessary, consider the limitations of relying solely on analytical physics. In theory, if you know the exact thermodynamic properties of a system, you can perfectly predict its thermal behavior. In practice, extreme environments and structural complexity make this impossible.

For example, in heavy industries like steel manufacturing, Tian et al. point out that classical thermodynamic equations for ladle furnaces require specific coefficients that cannot be measured due to the ultra-high-temperature environment. Similarly, Chen et al. note that physics-based carbon diffusion models for industrial heat treatment require precise interface transfer coefficients that vary unpredictably inside a real furnace.

Pure Physical Modeling (Fails in practice):
[Known Inputs] ---> [Physical Equations + UNKNOWN PARAMETERS] ---> [Inaccurate Predictions]

Hybrid Parameter Estimation (Succeeds):
[Known Inputs] ---> [Physical Equations + Latent Variables] ---> [Predicted Outputs]
                               ^                                      |
                               |------ [Optimization / ML Engine] <---| (Minimizes Error)
                                              ^
                                              |
                                      [Observable Data]

This problem is equally acute in advanced electronics. In electric motors, Lumped-Parameter Thermal Networks (LPTNs) offer a computationally efficient way to monitor internal heat. However, as Phuc et al. demonstrate, determining the internal thermal parameters of these networks analytically is notoriously difficult because the internal components are physically enclosed and inaccessible to standard sensors. When a model's predictive accuracy is entirely dependent on parameters that cannot be verified, the model cannot be safely deployed for real-time control.

Inverse Analysis: Running Physics Backward

The foundational mechanism for solving this problem is inverse analysis. Instead of using known parameters to calculate an unknown output (forward modeling), inverse analysis takes the observed outputs and works backward through the governing physical laws to deduce what the internal parameters must be.

This approach transforms a measurement problem into an optimization problem. For instance, Ni et al. utilize inverse heat transfer analysis to continuously determine the temperature-dependent thermal conductivity of explosives during solidification. Rather than attempting to isolate and measure the material at every volatile phase, they observe external temperature drops and calculate the underlying conductivity profile that mathematically satisfies the heat equation.

When direct mathematical inversion is too complex, intelligent optimization algorithms step in to search the parameter space efficiently:

  • Expectation-Maximization: Ševčík et al. adapted the Expectation-Maximization algorithm for power modules, allowing them to identify internal thermal parameters even when temperature data from critical active components is entirely missing.
  • Genetic Algorithms: In complex solid-gas sorption systems where phase-change heat capacities are unobservable, Li et al. deploy multi-population genetic algorithms to prevent the parameter search from getting stuck in local optima.

Resolving Ambiguity via Complementary Parameter Fusion

A major challenge in inverse modeling is identifiability: if multiple different internal parameter values yield the exact same external observation, the model cannot determine the true state of the system. This is a well-known issue when trying to estimate the junction temperature of Silicon Carbide (SiC) MOSFETs using a single Temperature-Sensitive Electrical Parameter (TSEP).

To resolve this ambiguity, Qin et al. introduced a method that leverages the rich thermal information embedded in the transient turn-off phase of the semiconductor. Instead of relying on a single parameter, the authors combine two distinct, complementary parameters: current fall time ($t_{fi}$) and current fall energy loss ($E_{fi}$).

$$\text{Junction Temperature } (T_j) = f(t_{fi}, E_{fi})$$

Because these two parameters respond to internal temperature variations in fundamentally different ways, combining them via multivariate linear regression isolates a unique thermal signature, drastically improving prediction accuracy where single-parameter methods fail.

Physics-Informed Machine Learning and Hybrid Modeling

The most recent evolution in this field moves away from pure optimization and toward tight integration with machine learning. Purely data-driven models (like standard neural networks) can map inputs to outputs easily, but they require massive datasets and often make predictions that violate physical laws. Hybrid modeling uses physics as a regularizer, forcing the machine learning model to operate within realistic boundaries.

In the domain of material science, empirical models like the Slack model are used to predict lattice thermal conductivity, but they require hard-to-obtain inputs like sound velocity. To bypass this, Liu et al. developed the PINK framework, which couples a crystal graph convolutional neural network with a simplified physical model. The machine learning component extracts structural features directly from crystal files, while the physical model ensures the final thermal conductivity predictions remain physically sound.

Similarly, for safety-critical systems like electric vehicle batteries, Chen et al. combined data-driven K-Means clustering with the model-based Bernardi equation to provide early warnings for battery thermal runaway. By using data-driven methods to absorb the uncertainty of unmeasurable parameters like the entropy heat coefficient, they achieved a 25-minute advance warning window with minimal false alarms.

Why the Hybrid Approach Works

The reason this research direction is so effective across diverse fields—from steel manufacturing (Xing et al.) to forensic thermal trace tracking (Xu et al.)—comes down to a fundamental division of labor:

  1. Physics provides the structure: Governing equations (like Newton's Law of Cooling or Fourier's Law) drastically reduce the search space. The machine learning or optimization algorithm does not have to learn the laws of thermodynamics from scratch.
  2. Data absorbs the real-world messiness: Optimization algorithms and regression models act as a sponge for the unmeasurable variables, adjusting the latent parameters until the physical model's output perfectly aligns with reality.

By shifting the goal from physically measuring a parameter to computationally isolating it, this paradigm allows engineers to deploy highly accurate, real-time physical models in environments that were previously considered black boxes.

Key Foundational Work

Hybrid Modeling of Molten Steel Temperature Prediction in LF (2008) — contributed 1 concept elements to this direction.
Hybrid intelligent parameter estimation based on grey case-based reasoning for laminar cooling process (2012) — 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_physical_parameter_measurement_gap_bd4ce180,
  title = {physical parameter measurement gap},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/physical-parameter-measurement-gap/rdr_bd4ce180713043b288d8cd5c3f4abbeb},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_physical_parameter_measurement_gap_bd4ce180_v1,
  title = {physical parameter measurement gap},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/physical-parameter-measurement-gap/rdr_bd4ce180713043b288d8cd5c3f4abbeb/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake