Parallelization and Optimization of KS Function Computation in Fortran for Aerodynamic Shape Optimization

This report presents the parallelization and optimization of the Kreisselmeier-Steinhauser (KS) function computation for aerodynamic shape optimization, implemented in Fortran and benchmarked against a TensorFlow-based version. The problem is distributed over multiple processors using an owner-computes approach for both the KS value and its derivatives, utilizing MPfor communication. Key optimizations include the use of a bounding box test to eliminate unnecessary triangle interactions and dynamic load balancing to maintain high processor utilization. These improvements resulted in significant speedups, achieving over 500x faster derivative computation compared to the TensorFlow implementation, with strong and weak scaling efficiencies explored in detail. Further optimizations are proposed to reduce collective communication and improve scalability for larger problems.