Enhancing Performance of Machine Learning-Based Modeling of Electromagnetic Structures

The machine learning (ML)-based modeling of electromagnetic (EM) structures involves the development of a surrogate model that approximates the relationship between EM geometries and responses, such as S 11 , gain, etc. The performance of the surrogate model is mainly affected by the simulation data for training. Normally, the training data is collected by uniformly sweeping the geometric parameters. Restricted by the computation power, only a limited parameter space can be sampled. The trained surrogate model behaves well within the sampling range but deteriorates as the parameter range extends. In this paper, we expand the predictable parameter range of an ML model with the same simulation expense by optimizing the data acquisition strategy. This approach leads to the proposed model demonstrating higher accuracy within an extended parameter space than conventional models, while the simulation consumption remains the same. We present an application example to validate its effectiveness. The proposed modified ML-based design method can potentially improve the performance of surrogate models in real-world applications.