研究目的
To increase the speed of gradient based EM optimization when no coarse model (e.g., empirical or equivalent circuits) is available by using a large number of fine model evaluations to achieve an overall speedup.
研究成果
The proposed trust-region based optimization technique achieves speedup in the optimization process for microwave structures using parallel computational approach, even when a coarse model is not available. By using large and effective optimization updates in each iteration, the method results in fewer optimization iterations. The technique demonstrates high parallel efficiency and robustness with respect to different starting points, making EM optimization more practical for microwave circuit design.
研究不足
The technique requires a large number of fine model evaluations, which may not be feasible for all designs due to computational resource constraints. The method's efficiency depends on the availability of parallel computational resources. Additionally, the technique assumes the surrogate model can accurately represent the EM behavior across the entire trust region, which may not always be the case for highly nonlinear or complex structures.
1：Experimental Design and Method Selection:
The proposed technique uses a large number of fine model evaluations to build a surrogate model valid in a large neighborhood. These valid surrogate models are used to achieve large and effective optimization updates, resulting in fewer iterations of the optimization process. Parallel computation is employed to reduce the computational time required for constructing a surrogate model. Trust region algorithms are exploited to guarantee convergence and to re-define the fine model evaluation range in each iteration.
2：Sample Selection and Data Sources:
Multiple sample points are generated around a central point using design of experiments (DOE) sampling strategy, specifically orthogonal sampling, to ensure the surrogate model is valid in a relatively large neighborhood.
3：List of Experimental Equipment and Materials:
A hybrid distributed-shared memory computational architecture, specifically a cluster of Dell PowerEdge computers with Intel Xeon X5680 processors, is used for parallel computation of fine model evaluations.
4：Experimental Procedures and Operational Workflow:
Fine model evaluations are distributed across all available computers to balance the workload. Each processor executes a job independently, and the results are collected to construct the surrogate model. The surrogate model is then optimized using trust region framework to determine the next optimization update.
5：Data Analysis Methods:
The surrogate model is constructed using transfer functions in the rational function format, and vector fitting method is used to extract coefficients. Regression techniques are applied to estimate the weighting vectors for all coefficients, and the surrogate model is optimized to find the next prospective central point.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
