Date of Award:


Document Type:


Degree Name:

Master of Science (MS)


Computer Science

Committee Chair(s)

Nicholas Flann


Nicholas Flann


Mario Harper


John Edwards


An urgent need is to electrify transportation to lower carbon emissions into the atmosphere. Wireless charging makes electrical vehicles (EVs) more convenient and cheaper because energy is transferred to the vehicle without the need to plug it in. Dynamic wireless charging is particularly interesting, where the vehicle does not need to stop to receive the energy. This technology requires the EV and the roadway to include coils of wire, where the roadway coil is energized as the vehicle passes over it to induce an electrical current in the EV coil through electromagnetic induction. However, the problem of designing the two coils (EV and road) is complex due to the many configurations possible, the need to maximize power transfer, and the need to minimize stray, possibly dangerous, electromagnetic fields during operation.

Current methods for designing inductive power transfer (IPT) coils rely heavily on FEM (finite-element methods) simulations to evaluate each potential design. Dynamic IPT design requires multiple simulation runs as the EV coil passes over the roadway coil. Identifying optimal designs is difficult because of the many conflicting specifications and objective functions that need to be considered, such as maximizing the output power while minimizing stray magnetic fields and the volume of windings and magnetic cores.

This work introduces a new design optimization method for dynamic IPT systems that utilize generative neural networks. Deep learning is applied to create a generator of near-optimal design alternatives from random noise. Two neural networks are employed in the approach. The first neural network is trained from multiple FEM results through random sampling of the design space and then replaces FEM calculations, allowing rapid simulation and evaluation of alternative designs. By using the neural network as a surrogate model rather than FEM to evaluate designs, differentiable programming approaches may be applied to train the second neural network to generate better designs. This generative network is trained by minimizing a loss function based on the optimization criteria listed earlier. Alternative loss function based on combining multi-objective optimization methods are explored, including applying the following mathematical operations over the objective functions: the sum of squares, a product of means, and sums of combinations of pair-wise products.

Compared to previous work [1], which employed genetic algorithm approaches, the generative network quickly learned to produce designs that pass all objective functions using the product of means, however, the design solutions lacked diversity. Interestingly, when considering all pairwise product combinations only a few worked in quickly learning to produce satisfactory solutions. Those combinations that worked had a lower solution production rate than the product of means but exhibited a higher diversity of solutions.