Using a Domain Theory to Guide Automated Modeling of Complex Physical Phenomena

Sui-ky Ringo Ling

Introduction

In engineering and scientific analysis, selecting the appropriate mathematical model is a high-stakes balancing act. A model that is too complex, such as a system of partial differential equations (PDEs), can consume excessive computational resources. Conversely, a model that is too simple may omit critical physical phenomena, leading to inaccurate results. This paper introduces MSG (Model Selection and Generation), a system designed to automate the creation of these models by leveraging structured domain theory.

The "Vicious Circle" of Modeling

The author identifies a fundamental challenge in automated modeling termed the "vicious circle": to choose an appropriate model, an engineer must already understand how the system behaves (e.g., determining if temperature is uniform across a block). However, determining that behavior typically requires a model in the first place. Traditionally, human intuition bridges this gap. MSG seeks to automate this intuition by using a structured domain theory to guide a computer through the modeling process.

Methodology: Structured Domain Theory and Estimation

The MSG system relies on two primary pillars to break the modeling circle:

  1. Structured Domain Theory: Rather than treating equations as arbitrary symbols, MSG organizes modeling around fundamental conservation laws, such as the conservation of energy. It views a model as a composition of basic physical processes (conduction, convection, radiation) acting on defined "control regions."
  2. Gross Behavior Estimation: MSG uses simple, approximate metrics to estimate a system's behavior before a full model is built. For instance, it calculates the Biot number ($hL/k$) to decide whether a solid can be treated as having a uniform temperature (a lumped-parameter model) or if it requires a distributed PDE model.

The modeling process follows a three-step sequence: defining control regions, identifying active physical processes based on threshold approximations, and transforming these choices into formal, well-formed mathematical equations including initial and boundary conditions.

Key Insights: Physics-Aware Simplification

A core innovation of MSG is its approach to simplification. Unlike purely mathematical or "syntactic" simplifications that might strip an equation of its physical meaning, MSG ensures that approximations remain consistent with domain theory.

Illustration of physics-aware simplification for radiation heat transfer.

As shown in the figure above, when $T_1 \gg T_2$, a standard mathematical simplification might reduce the radiation term to $\sigma \epsilon T_1^4$. However, MSG recognizes that radiation, like convection, fundamentally depends on a temperature difference. The system ensures that approximations occur by eliminating entire physical processes rather than just symbols, keeping the resulting model physically grounded.

Results and Performance

MSG was evaluated against 27 problems from standard engineering textbooks and 16 hand-crafted edge cases involving complex geometries like L-shaped blocks. The results demonstrate significant efficiency and accuracy:

  • Efficiency: MSG can formulate models involving coupled PDEs in less than 4 seconds on a SparcStation 2.
  • Accuracy: The system is calibrated to produce models with a typical error of less than 10%, meeting standard engineering requirements.
  • Versatility: The system successfully automates the transition from a physical description (geometry and materials) to a complete mathematical model.
A complete mathematical model generated by MSG for a composite plate, including governing equations, initial conditions, and boundary conditions.

The figure above illustrates the output of MSG: a rigorous set of governing equations and boundary conditions that define the thermal behavior of a composite system. By sequencing decisions from control regions to processes to equations, MSG drastically reduces the search space, allowing for rapid model generation.

Conclusion and Future Work

MSG demonstrates that a domain-theory-driven approach can effectively automate complex modeling tasks. While currently optimized for well-established fields like heat transfer and restricted to regular geometries, future work aims to expand the system to arbitrary 3D geometries and develop dynamic methods for calibrating approximation thresholds based on user-defined accuracy needs.

Summary updated

BibTeX
@inproceedings{ling1995using,
  author = {Sui-ky Ringo Ling},
  title = {{Using a Domain Theory to Guide Automated Modeling of Complex Physical Phenomena}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1995},
}
Figure