Inference of electromagnetic system behavior in the presence of variability

Stochastic electromagnetic-circuit analysis;Statistical inference;Generative modeling;S-parameters;Variational autoencoder;Physical constraints;Passivity;Dependency between subsystems;Conditional variational autoencoder;Universal statistics;Random coupling model;Wave chaos

Ma, Xiao

关键词: Stochastic electromagnetic-circuit analysis;    Statistical inference;    Generative modeling;    S-parameters;    Variational autoencoder;    Physical constraints;    Passivity;    Dependency between subsystems;    Conditional variational autoencoder;    Universal statistics;    Random coupling model;    Wave chaos;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/107854/MA-DISSERTATION-2020.pdf?sequence=1&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

This thesis contributes to and advances the state-of-the-art of the analysis of stochastic electromagnetic-circuit systems by applying novel statistical inference and generative modeling techniques to the extraction of statistical models of system responses. Machine learning based and physics based techniques are introduced to address the limitations of existing methods by incorporating dimensionality reduction into the model training procedure, enabling the enforcement of certain physical constraints, making the modeling procedure more flexible by accepting additional data to capture dependency, and taking advantage of universal statistics to reduce the sample requirement.Specifically, for the first time, a variational autoencoder based method is used for the generative modeling of high-dimensional S-parameters data. The generation accuracy is shown to be superior to that of existing methods.The passive variational autoencoder, a variational autoencoder with modified decoder architecture, is introduced to enforce physical constraints like passivity. The generated S-parameters are shown to be physically consistent. The generation accuracy and efficiency of the proposed method are assessed through comparison to those of existing methods.When a system is decomposed into multiple subsystems, the dependency between responses of subsystems needs to be captured but is often neglected. A conditional variational autoencoder is introduced for the capturing of such dependency. Additional control variables are used to account for the randomness external to a subsystem. The conditional variational autoencoder based generative model demonstrates an advantage in terms of generation accuracy as compared to its standard version, which ignores the dependency between subsystems.The random coupling model is a physics based generative model applicable to electromagnetic enclosures under wave chaos condition. The low-frequency limit of the random coupling model based generative model is shown to be assessed by computing the random projection Kolmogorov-Smirnov statistics, and the low-frequency limit is demonstrated to be lower than the rule-of-thumb estimate used in past works.