Training Set Optimization with Uncertainty Quantification for Machine Learning Models of Electromagnetic Structures

Neural Networks surrogate modeling for EM simulations saves computational and design time. Introducing uncertainty estimates into deterministic prediction models provides insight into the reliability and confidence of the model. However, gathering training data to train models is a very time-consuming and resource-consuming task. In this paper, we introduce a method to harness useful insights from confidence bounds to reduce the training set size required to train a model with reasonable accuracy and latency. Using a high-speed differential via structure, we show that the training samples required are 35% less with a slight trade-off in accuracy using the proposed method.