Y.V. Kuznetsov, G.P. Sinyavsky

Space mapping (SM) optimization methods are special case of optimization techniques that use surrogate models. In general, SM surrogate includes coarse model of the device, an input mapping to transform input surrogate model parameters into coarse model parameters, and an output mapping to ensure consistency. SM surrogate needs a few fine model evaluations to build, and also has good predictive properties. As a consequence, optimization time is drastically reduced. In order to avoid overfitting and facilitate parameter extraction, input and output mappings need not to be too flexible. Our experience and a review of literature covering the subject, both show that the simple, yet effective surrogate model, that is suitable for most practical cases, should employ linear input mapping and additive output mapping. We suggest finding input linear mapping using space derivative mapping method and modified linear inverse space mapping method. Linear inverse space mapping also simplifies obtaining the optimum change of fine model parameters. If equivalent circuit are used as a coarse model, fine model parameters and coarse model parameters have different physical meaning. It is more is convenient then to use indirectly found geometric parameters as coarse model parameters instead. Indirect coarse model parameters can be found using analytic functions which relate electric circuit parameters and corresponding geometric parameters of the device. These functions usually are known, because they are used to synthetize initial geometry of the device. The new approach facilitates use of equivalent circuits: initial approximate of input mapping is considered to be unity mapping, and additional upfront effort needed to evaluate link between geomitric and electric parameters is avoided.