Expedited Feature-Based Local Optimization: Adjoint Sensitivities and Sparse Sensitivity Updates

The main contributor to the computational cost of solving local optimization tasks using gradient-based routines with numerical derivatives is the estimation of the response sensitivity through finite differentiation. Chapter 7 discussed procedures developed to reduce this cost by employing variable-fidelity EM models. This chapter focuses on alternative techniques that enable fast evaluation of the system response gradients. The considered methods may be incorporated into feature-based optimization routines. Two specific acceleration mechanisms are highlighted: adjoint sensitivities and sparse sensitivity updates. The first technique may be considered less versatile, as it is intrusive from the point of view of the EM simulation engine, and it is only accessible through a few commercial simulation packages. The second option is sparse sensitivity updates where costly finite-differentiation-based sensitivity updates throughout the algorithm run are suppressed to a certain degree. Several different mechanisms for restraining the updates have been developed. These include analysis of the design relocation between consecutive iterations, monitoring of the magnitude of gradient changes during the optimization run, and restricting updates to the directions of the most significant changes of the structure responses. Another approach involves the employment of updating formulas. In this chapter, FBO procedure is expedited by detecting patterns of response gradient variability throughout the algorithm run, and suppressing sensitivity updates for the variables that exhibit stable gradient behavior.