Microextrusion Shape Optimization

About this Project

For my final project in Dr. Dan Stoecklein’s Design Optimization elective, I contributed to his work developing a design tool for microextrusion die shapes. I started with his tool that quickly calculates the final shape of a microextrusion given a die shape, then formulated it as a 10-D standard optimization problem. Then, I compared the performance, number of evaluations, and wall time of several common optimization algorithms.

Optimization Algorithms Compared

• Original Problem: Each optimization algorithm was asked to maximize the similarity of an output image to a specified target shape. Initial conditions were either a circular die or a slightly perturbed version of the “solution” die.
• Nelder-Mead: Performed the best of all three algorithms, finding satisfactory visual matches and high similarity with only ~500 function iterations.
• Particle Swarm Optimization: Used 3000-4000 function evaluations, but was able to find highly performant solutions.
• Steepest Descent: Achieved the highest similarity score in most cases, but required upwards of 7000 function evaluations to do so.