Professor Hod Lipson, Sofia Wyetzner

Creative Machines Lab, Columbia University

June - December 2019

GE Jet Engine Bracket optimization.

I was assigned to test and debug a program being developed by the Creative Machines Lab at Columbia University, the Design Model Language interactive development environment.

My role as an undergraduate researcher was to test the program by creating different structures to simulate and to find areas in the code and runtime that could be improved.

I used CAD programs, C++/C++ with CUDA and the Nvidia Profiler to complete these tasks.

Torsional beam optimization.
The program, developed in C++, is used to render 3D structures as mass-spring objects and simulate forces on it. The program then optimizes the mass of the structure by removing springs under the least amount of stress. The program halts when the object reaches a certain deflection rate or percentage of original mass remaining.
The Interactive Development Environment was used to simulate any imported volume under a specified force. This volume was the result of a cantilever beam subject to a torsional force on one end. The result is a hollowed rectangular prism of springs with a spiraling effect. My role was to ensure that the program was able to handle different cases and simulations, making it more applicable to different real-world uses.
Nvidia Visual profiler example.

Because the program was very computation heavy when importing large objects that would turn into hundreds of thousands of mass-spring objects, we wanted to benchmark the program in different configurations to find the fastest configuration of memory organization and thread-block use.

Using the Nvidia Visual Profiler, we were able to discern which kernels hosted the longest computation times, and needed reducing. We found that the function with the highest usage of processing power was a function called computeSpringForces( ), which is a necessary step in each timestep of the simulation.

Nvidia visual profiler.
Bracket before optimization.
Bracket after optimization.