A Machine Learning Based Approach for Post Manufacturing Performance Sensing in Analog Circuits

As CMOS technology advances towards miniaturization for enhanced performance, the impact of process uncertainties on analog circuits becomes increasingly pronounced, often resulting in degraded performance. In order to ensure proper functionality, the fabricated analog circuits are subjected to performance check during the production test where these are interfaced with high end Automated Test Equipment (ATE), leading to higher test cost. This work introduces a low-cost Machine Learning (ML) based non-intrusive sensing strategy for performance estimation of process perturbed analog circuits. The sensor outputs, termed as signatures are used for performance prediction using a pre-trained ML model. These sensors are placed close to the Circuit Under Test (CUT), provides an image of process variations, without actually loading it. A systematic approach for identification of sensor signatures is proposed using Recursive Feature Elimination Technique. The proposed performance sensing approach is validated on a current-starved voltage-controlled oscillator (CSVCO), where a neural network-based regression model predicts the key performance metrics in a CSVCO.