The deeper subsurface layers beyond the lunar regolith are not well-constrained. Using in-situ lunar penetrating radar data (d) from Change’E-3 Yutu rover, we invert for its subsurface relative dielectric permittivity (εr) model. We use a hybrid physics informed convolutional neural network based deep learning architecture to simultaneously regress the relationship between εr – d; and using radar forward modeling to reconstruct radargrams to match the field data d. This dual pronged inversion approach will ensure predicted radargrams to closely match the field data and providing reasonable εr estimates. Our training dataset will incorporate a priori knowledge of existing εr ranges [1-10] from previous lunar regolith studies. We expect to predict accurate εr profiles of the lunar subsurface (~200 meters) and generating predicted data that closely matches the field data.