Collaborative Research: Planning Grant: I/UCRC for Advanced Electronics through Machine Learning

The accepted engineering design methodology requires that mass scale manufacturing of a new product not commence until a prototype of the product is tested and found to meet its performance specifications. It is not unusual for a product to go through multiple design iterations before it can satisfy all the design requirements. Modern electronic products, which range from a single integrated circuit to a smart phone to an aircraft instrumentation system, are so complex and contain so many components - billions in the case of an integrated circuit - that it is infeasible to construct hardware prototypes for each design iteration, from the points of view of both cost and time. Instead, a mathematical representation of the product must be developed, i.e. a virtual prototype, and its behavior then simulated. Each of the components that constitute the product would be represented by a model. Behavioral models of the components are most desirable; a behavioral model represents the terminal response of a component in response to an outside stimulus or signal, without concern to the inner workings of the component. Behavioral models are computationally efficient and have the benefit of obscuring intellectual property. However, despite many years of significant effort by the electronic design automation community, there is not a general, systematic method to generate accurate and comprehensive behavioral models, in part because of the non-linear, complex and multi-port nature of the components being modeled. The proposing team will utilize the planning grant to establish a research center that will overcome these modeling challenges through the development and application of novel machine-learning methods and algorithms.

Machine-learning algorithms are used to extract a model of a component or system from input-output data, despite the presence of uncertainty and noise. In this center, the input-output data are obtained either from measurements of a component or by running detailed simulations of a component. The emphasis is on models that balance good predictive ability against computational complexity. The center will pioneer the application of machine learning to electronics modeling. It will develop a methodology to use prior knowledge, i.e., physical constraints and domain knowledge provided by designers, to speed up the learning process. Novel methods of incorporating component variability, including that due to semiconductor process variations, will be developed.