A Support Vector Regression based Machine Learning method for on-chip Aging Estimation

Semiconductor supply chain industry is spread worldwide to reduce cost and to meet the electronic systems high demand for ICs, and with the era of internet of things (IoT), the estimated numbers of electronic devices will rise over trillions. This drift in the semiconductor supply chain produces high volume of e-waste, which affects integrated circuits (ICs) security and reliability through counterfeiting, i.e., recycled and remarked ICs. Utilising recycled IC as a new one or a remarked IC to upgrade its level into critical infrastructure such as defence or medical electronics may cause systems failure, compromising human lives and financial loss. This paper harvests aging degradation induced by BTI and HCI, observing frequency and discharge time affected by changes in drain current and sub-threshold leakage current over time, respectively. Such task is undertaken by Cadence simulations, implementing a 51-stage ring oscillator (51-RO) using 22nm CMOS technology library and aging model provided by GlobalFoundries (GF). Machine learning (ML) algorithm of support vector regression (SVR) is adapted for this application, using a training process that involves operating temperature, discharge time, frequency, and aging time. The data sampling is performed over an emulated 12 years period with four representative temperatures of 20° C, 40° C, 60° C, and 80° C with additional testing data from temperatures of 25° C and 50° C. The results demonstrate a high accuracy on aging estimation by SVR, reported as a normal distribution with the mean (µ) equal to 0.01 years (3.6 days) and a standard deviation (σ) of ±0.1 years (±36 days).