Batch Training of Gaussian Process for Up-sampling Problems in S-Parameter Predictions

In using Machine Learning (ML) methods to predict S-parameters, handling the dimensionality problem of mapping the low-dimension design parameters to high-dimension responses is important. We propose to use batch training of Gaussian Process (GP) to map the design parameters into latent Gaussian space instead of linear mappings to create the non-linearity property as well as avoiding the saturation of activation functions before applying transposed kernels. Results show that the proposed model achieves better performance with regard to loss and normalized mean-squared error.