Date of Award:


Document Type:


Degree Name:

Master of Science (MS)


Computer Science

Committee Chair(s)

Nicholas Flann


Nicholas Flann


Yuan Shuhan


Vladimir Kulyukin


Electric vehicles (EVs) offer many improvements over traditional combustion engines including increasing efficiency, while decreasing cost of operation and emissions. There is a need for the development of cheap and efficient charging systems for the future success of EVs. Most EVs currently utilize static plug-in charging systems. An alternative charging method of significant interest is dynamic inductive power transfer systems (DIPT). These systems utilize two coils, one placed in the vehicle and one in the roadway to wirelessly charge the vehicle as it passes over. This method removes the current limitations on EVs where they must stop and statically charge for a period of time. However, the physical designs of the charging unit coils depends on many physical parameters, which leads to complexity when determining how to design the unit. A design then needs to be judged for its quality and performance, for which there are eight proposed objective functions. These objective functions represent performance metrics but are conflicting. Some metrics, such as output power are to be maximized, while others such as stray magnetic field and volume of windings and magnetic cores are to be minimized. Different unit designs will trade off performance for these objectives.

In order to address the complex issue of finding near-optimal designs, a machine-learning, Generative Neural Network (GNN) approach is presented for the rapid development of near-optimal DIPT systems. GNNs generate new examples from a trained neural network and have demonstrated remarkable power in creating novel graphic design images from text-to-image training. This stems from their ability to be creative yet constrained in a regular domain. In this case, a simulator network and evaluator generator network are implemented. The simulator is a pre-trained neural network that maps from the physical designs to the objective functions. The generator network is trained to generate the near-optimal physical designs. A design is considered successful if it passes given thresholds for each of the eight objective functions, which evaluate the quality of a design. Before training, the rate of finding successful designs is 0.005%, but within 500 training epochs the rate becomes 98% (about 30 seconds total GPU run time). Engineers and production professionals are interested in both the best performing designs as well as a diversity of configurations to build. In order to improve on these criteria, several different loss functions were developed that incorporate the objective functions. Loss functions are what the neural networks use to determine how to adjust parameters and produce a better design. The various loss functions presented greatly influence the ability of the GNN to produce diverse and high-performance design solutions.