A major part of the design process is testing out numerical codes. The numerical codes can either be "home grown" -- i.e. developed by the designer, or acquired from other scientists. Numerical codes are complex, expensive and unreliable. When a code is tested or used for design purposes, part of this process includes determining the limits of the underlying mathematical models.

Problems

Engineers commonly ask some of the following questions when they are trying optimize design based upon a numerical code. All of these questions involve looking at the code and trying to analyze the behavior of the code and the behavior of the mathematical model.

• How good is the model?
• Over what regions is the model valid?
• Where does failure occur?
• What parameter space effects the model?

• What physical factors are contributing to the design process?
• Can these factors be broken up into parts to see the influence of different parameter changes vs outcome changes?

• Is the model at fault or the code logic?
• Can a different solution be plugged into the existing code?

• What does my resulting dataset look like?

Approaches

In this project, we propose to develop a set of visualization routines which will help the scientist assimilate the programs and model.

• Use computational experiments and visualization techniques to understand behavior of simulation model and interaction between optimizer and simulator.
• Identify regions of input space that violate qualitative expectations, or lead to crashing.
• Identify causes of such violations or crashing.
• Identify pathologies: Ridges, discontinuities, Multiple-Local Optima and causes of such pathologies.
• Diagnose failures of optimization process.
• Identify suitable approximations.

Preliminary Techniques

We use nozzle/aircraft domain as case study. We are currently developing visualization tools to help scientist to assimilate the programs and model.

A prototype of visualization environment has been developed to attack some of the problems in the optimization/simulation experiments

• Visualizing regions associated with each type of failure.

• Visualizing the take-off mass

• Visualizing search path taken by optimizer with different strategy:
• Strawman Strategy

• Two-Stage Strategy

• Visualizing local optimum points of different optimizations.

• Identify the cause inside the code that causes the type of error

By clicking on the error region, users are presented with the dataflow graph showing the piece of code that generated that type of error.

Expected Impact

• Dramatic improvement in the productivity of working scientist and engineers.
• More widespread disemmination of models.
• More widespread use of optimization/simulation.

Visit Vizlab

Return to HPCD Home Page