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

TL;DR

This paper introduces an approach to automatically generate mathematical models for complex physical phenomena by leveraging established "domain theories" and problem-specific behaviors. By structuring the modeling process around fundamental conservation laws and using simple approximations to understand a problem's gross behavior, this method offers a uniform and efficient way to create models of varying complexity, from simple algebraic equations to complex partial differential equations.

What Problem Does This Paper Address?

Developing accurate mathematical models for complex physical phenomena is a cornerstone of scientific computation and engineering analysis. However, this process is often challenging, especially when dealing with high-performance computing requirements and intricate partial differential equations. The core problem this paper addresses is how to automate the formulation of these models in a way that is both efficient and produces results consistent with the underlying physics. Traditional approaches often require significant expert knowledge and manual intervention. The authors highlight that while many physical domains possess rich "domain theories" (fundamental laws and well-defined processes), effectively translating these theories into computational models, especially when approximations are necessary, remains a significant hurdle. The challenge lies in making informed modeling decisions that balance accuracy, computational cost, and adherence to physical principles.

Methods

The paper proposes a two-principle approach to automated modeling:

  1. Leveraging Domain Theory Structure: The modeling process is organized around the fundamental conservation laws of a domain. This provides a template for composing physical processes. Approximations are then focused on selecting the types and number of these basic physical processes.
  2. Using Gross Behavior for Guidance: Simple, approximate models are used to estimate the overall behavior of a physical problem. This "gross behavior" (e.g., maximum temperature drop rather than a full distribution) guides the modeling decisions, helping to determine appropriate control regions and the complexity of physical processes to include.

This approach has been implemented in a computer program called MSG (Model Selection and Generation), specifically for the heat transfer domain. MSG follows a structured process:

  • Preprocessing: Identifies common surfaces and computes thermal resistances.
  • Choices of Control Regions: Determines whether to use lumped or differential control regions based on thermal resistance ratios and a "lumping threshold."
  • Choices of Physical Processes: Selects relevant heat transfer processes (conduction, convection, radiation) and applies approximations based on transient and spatial thresholds.
  • Mathematical Transformations: Translates the chosen processes and control regions into mathematical equations, including initial and boundary conditions.

Datasets

Not discussed in the paper. The paper mentions testing MSG with 27 problems from two heat transfer textbooks and 16 hand-crafted problems.

Key Insights and Ideas

  • Domain Theory as a Blueprint: The paper argues that a well-established domain theory, with its conservation laws and defined physical processes, provides a powerful structural framework for building models. This is more effective than purely compositional approaches that might lack this underlying physical grounding.
  • Gross Behavior as a Compass: Using simple, approximate models to understand a problem's overall behavior is a clever way to break the "model-requires-knowledge, knowledge-requires-model" circular dependency. This allows for informed decisions about model complexity without needing a fully formed model upfront.
  • Compositionality at the Process Level: The approach emphasizes that models are built by composing physical processes, rather than just assembling pre-defined model fragments. This allows for more flexibility and a more direct link to the underlying physics.
  • Consistency Over Pure Approximation: The authors stress the importance of ensuring that approximations lead to models consistent with the domain theory. They illustrate how naive application of approximation rules can lead to physically nonsensical models, and their approach aims to prevent this by constraining approximations to the selection of physical processes.

Key Results

The MSG program successfully formulated mathematical models for various heat transfer problems, ranging from simple algebraic equations to coupled nonlinear partial differential equations. The models generated were consistent with the heat transfer domain theory. The paper states that MSG can formulate a model involving coupled partial differential equations in less than 4 seconds on a SparcStation 2, demonstrating significant efficiency. The approach was tested on problems involving energy storage, heat flux, and temperature distribution, covering single and multiple components in rectangular geometries.

Limitations and Future Work

The authors acknowledge that this approach is heavily reliant on the existence of a strong domain theory with well-defined conservation laws and approximate numerical methods. This limits its applicability to well-established physical domains. Future work could explore extending this approach to other domains with similar theoretical foundations, such as fluid mechanics. The paper also implicitly suggests that further calibration of thresholds for approximations could enhance the accuracy and robustness of the generated models.

Summary updated

BibTeX
@inproceedings{1995using,
  title = {{Using a Domain Theory to Guide Automated Modeling of Complex Physical Phenomena}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1995},
}
Figure