A Machine Learning Based Built-In Self-Test Method for Lifetime Frequency Drift Compensation of Precision Oscillators

Embedded RC oscillators are widely used as precision reference clock generators in modern system-on-chip (SoC) designs. The lifetime frequency performance of the RC oscillators is impacted by the aging degradation of the on-chip resistance. This paper presents a machine learning (ML) based method that estimates and compensates for the lifetime frequency drift. An ML based regression model that establishes a correlation between a set of static node voltages in the oscillator and its frequency is developed. The drift is proposed to be evaluated using onchip monitoring circuits, analog to digital converter and a neural inference engine integrated in the SoC, and the calibration settings are adjusted to maintain the lifetime frequency precision. The method is validated with simulations using an oscillator circuit with a nominal frequency of 10 MHz, where the model is tested under arbitrary conditions of aging, supply voltage and ambient temperature variations, and the maximum error in estimated frequency drift is within 2000 ppm.