This paper proposes a novel class of neural-network inspired statistical data-driven models, especially derived for the purpose of design optimization of medium frequency transformers. These models allow for an efficient (3-5 orders of magnitude faster compared to FEM), yet sufficiently accurate (within 5-10 % error relative to FEM) and numerically stable estimation of the complex effects, with otherwise impractically high computational cost and/or convergence issues. The application of the proposed modeling framework is described in detail on two characteristic examples of the complex electromagnetic phenomena occurring within the medium frequency transformers. The performance of the derived models is verified both with detailed FEM simulations and experimental results.