Automated Physical Modeling

Andrew GelseyIJCAI 1989

Introduction

In the late 1980s, Artificial Intelligence research into "qualitative physics" faced a significant hurdle: mechanical systems are defined by their precise geometry. While humans can intuitively grasp how a clock "ticks" by looking at a diagram, automating this understanding for a computer is remarkably difficult. Engineers typically construct physical models by hand, deciding which factors to include and which to neglect. Andrew Gelsey's work addresses this gap by presenting a system that automatically transforms 3D geometric descriptions of mechanical devices into functional quantitative models.

The Automated Modeling Pipeline

The system utilizes a multi-stage pipeline to derive the differential equations and state variables necessary for simulation.

1. Geometric Input and Kinematic Analysis

The process begins with a Constructive Solid Geometry (CSG) representation, a standard CAD format using basic shapes and boolean operations. The program then performs a kinematic analysis to identify "kinematic pairs" (e.g., a piston within a cylinder) and "chains" (e.g., a gear train). By identifying these relationships, the system can significantly reduce the number of state variables; for instance, an entire gear train can be represented by a single position and velocity variable rather than tracking every individual gear.

2. Force and Collision Modeling

To handle intermittent contact between parts, the system employs an overlap routine. While bodies are treated as rigid, the model allows for microscopic overlaps during collisions. This overlap generates a restorative force modeled as a damped spring:

$$f = E \cdot o + D \cdot \dot{o} \cdot M \cdot o$$

This approach creates a mathematical "recipe" for calculating time derivatives during simulation.

3D wireframe model of a clock escapement showing the balance, lever, and escape wheel components.

3. Behavior Prediction via Loop Jumping

Simulating every millisecond of a complex machine's operation is computationally prohibitive. Gelsey introduced a "loop-jumping" algorithm that monitors the system's kinetic energy. By identifying repeating patterns of energy minima, the program detects stable behavioral "loops." Once a loop is confirmed, the simulation can "jump" forward in time, predicting long-term behavior without brute-force calculation.

Key Results and Insights

The system was evaluated on classic mechanical benchmarks, most notably a clock/watch escapement. This mechanism is notoriously difficult to model due to its "stop-and-go" nature and frequent collisions.

Simulation plots showing the periodic oscillations of the balance and lever, and the stepped progression of the escape wheel.

As shown in the simulation results, the system successfully generated ordinary differential equations (ODEs) for the escapement. The balance wheel exhibits sinusoidal oscillation, the lever moves in discrete pulses, and the escape wheel advances in a characteristic "stepped" manner. The loop-jumping algorithm identified a cycle containing two global kinetic energy minima, allowing for efficient long-term prediction.

The "Neglect" Principle

A core insight of this research is that modeling is as much about what to exclude as what to include. Gelsey demonstrates that his program can generate multiple valid models of the same device—for example, one treating parts as perfectly rigid and another allowing for elastic deformation—depending on the required level of detail.

Conclusion and Future Directions

Gelsey's work demonstrates that geometry is not just a visual representation but a quantitative foundation for physical reasoning. While the current system is specialized for mechanics and requires clean CSG input, it provides a blueprint for automated physical modeling. Future work involves extending these "recipes" to other domains like thermodynamics and fluid dynamics, and automating the logic used to choose the most appropriate model for a given task.

Summary updated

BibTeX
@inproceedings{gelsey1989automated,
  author = {Andrew Gelsey},
  title = {{Automated Physical Modeling}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1989},
}
Figure