I/UCRC: Center for Advanced Electronics through Machine Learning (CAEML)

The semiconductor industry is perennially one of America's top exporters. Worldwide semiconductor

sales for 2014 reached $335.8 billion, and the number of U.S. jobs in this sector was estimated to be

around 250,000 in 2013. More broadly, the U.S. tech industry, which depends on semiconductor

innovation to spur new products and applications, is itself estimated to represent no less than 5.7% of the

entire U.S. private sector workforce (at nearly 6.5 million jobs), and with a tech industry payroll of $654

billion in 2014, it accounted for over 11% of all U.S. private sector payroll. Yet despite its success, the

industry must continue to innovate if the U.S. is to retain global leadership in this highly competitive area.

The complexity of modern microelectronic products necessitates the use of computer tools to formulate

and verify product designs prior to manufacturing. When a product doesn't operate as intended or suffers

early failures, this can often be attributed to inadequacy of the models used during the design process. In

fact, the shortcomings of existing approaches for system component modeling have become a serious

impediment to continued innovation.

The Center for Advanced Electronics through Machine Learning (CAEML) proposes to create

machine-learning algorithms to derive models used for electronic design automation with the objective of

enabling fast, accurate design of microelectronic circuits and systems. Success will make it much easier

and cheaper to optimize a system design, allowing the industry to produce lower-power and lower-cost

electronic systems without sacrificing functionality. The eventual result will be significant growth in

capabilities that will drive innovation throughout the electronics industry, leading to new devices and

applications, continued entrepreneurial leadership, and economic growth.

While achieving those goals, CAEML will also focus on diversifying the undergraduate engineering

student body and improving the undergraduate experience. Students from groups traditionally

underrepresented in engineering will be targeted for recruitment as undergraduate research assistants.

Member companies will provide internships and mentors for participating students, and the diverse

graduate and undergraduate student researchers in CAEML will receive hands-on multidisciplinary

education. CAEML will also participate in all three site universities' existing avenues for student and

faculty engagement with local youth. In particular, university-based summer camps are a tried and tested

method of making high-school students familiar with and comfortable on our campuses. The HOT DAYS

@ Georgia Tech (https://www.ece.gatech.edu/outreach/hot-days) camp is a week-long summer program

designed to introduce high-school students to electrical and computer engineering concepts through

various half-day modules including building a computer, working with robots, using music synthesis

technology, building simple digital logic circuits and constructing a speaker from common household

items. Additional modules covering CAEML research areas will be developed and incorporated into the

camp's schedule. CAEML undergraduate and graduate students can serve as counselors or instructors

for camps. Georgia Tech's Center for Education Integrating Science, Mathematics, and Computing

(CEISMC) hosts a variety of camps and programs for K-12 teachers, as well as students. CAEML faculty

at Georgia Tech will participate in that effort as well as the NSF-funded Summer Teacher Experience in

Packaging, Utilizing Physics (STEP-UP) Program (https://www.ece.gatech.edu/outreach/step-up-

program), which is an eight-week research experience for metro Atlanta high-school physics teachers.

The Center for Advanced Electronics through Machine Learning (CAEML) will create machine-

learning algorithms to derive models used for electronic design automation, with the objective of enabling

fast, accurate design of microelectronic circuits and systems. The electronics industry's continued ability

to innovate requires the creation of optimization methodologies that result in low-power integrated

systems that meet performance specifications, despite being composed of components whose

characteristics exhibit variability and that operate in different physical or signal domains. Today,

shortcomings in accuracy and comprehensiveness of component-level behavioral models impede the

advancement of computer-aided electronic system design optimization. The model accuracy also impacts

system verification. Ultimately, the proper functionality of an electronic system is verified through testing

of a representative sample. However, modern electronic systems are so complex that it is unthinkable to

bring one to the manufacturing stage without first verifying its operation using simulation. Today,

simulation generally does not ensure that an integrated circuit or electronic system will pass qualification

testing the first time, and failures are often attributed to insufficiency of the simulation models. With an

improved modeling capability, one could achieve better design efficiency, and also perform design

optimization. For system simulation, behavioral models of the components' terminal responses are

desired for both computational tractability and protection of intellectual property. 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 nonlinear,

complex, and multi-port nature of the components being modeled.

CAEML will pioneer the use of machine-learning methods to extract behavioral models of electronic

components and subsystems from simulation waveforms and/or measurement data. The Center will make

2 primary contributions to the field of machine learning: it will demonstrate the application of machine

learning to electronics modeling, and develop the entire machine-learning pipeline. Historically, machine-

learning theorists have focused on the model learning and evaluation tasks, but CAEML will focus on

end-to-end performance of the pipeline, including data acquisition, selection and filtering, as well as cost

function specification. CAEML will develop a methodology to use prior knowledge, i.e., physical

constraints and the 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. The intended end-users are electronic design automation (EDA) tool developers, IC

design houses, and system design and manufacturing companies.

CAEML consists of 3 sites: Illinois, Georgia Tech, and NC State. The scope of research at each site

encompasses both algorithm development and the application of the derived models to a variety of IC

and system design tasks. Investigators at all 3 university sites have unique skills and expertise while

sharing interests in electronic design automation, IC design, system-level signal integrity, and power

distribution. To leverage the cross-campus expertise, many of the Center's proposed projects involve

investigators from more than one site. The Georgia Tech investigators have special expertise in

advanced IC packaging, power integrity, multi-physics simulation, computational electromagnetics, neural

networks, optimization and system integration. All three sites have strong research records in the fields of

signal integrity analysis and electronic design automation. Excellent computational resources are

available at Georgia Tech for the proposed work including extensive measurement and fabrication

facilities.