On the Limits of Machine Learning-Based Test: A Calibrated Mixed-Signal System Case Study
Manuel J. Barragan1, G. Leger2,a, A. Gines2, E. Peralias2 and A. Rueda2
1CNRS, TIMA, F-38000 Grenoble, France, Université Grenoble Alpes, TIMA, F-38000 Grenoble, France.
manuel.barragan@imag.fr
2Instituto de Microlectrónica de Sevilla, CSIC-Universidad de Sevilla, Av. Américo Vespucio s/n, 41092 Sevilla, Spain.
aleger@imse-cnm.csic.es
ABSTRACT
Testing analog, mixed-signal and RF circuits represents the main cost component for testing complex SoCs. A promising solution to alleviate this cost is the machine learningbased test strategy. These test techniques are an indirect test approach that replaces costly specification measurements by simpler signatures. Machine learning algorithms are used to map these signatures to the performance parameters. Although this approach has a number of undoubtable advantages, it also opens new issues that have to be addressed before it can be widely adopted by the industry. In this paper we present a machine learning-based test for a complex mixed-signal system 215 i.e. a state-of-the-art pipeline ADC 215 that includes digital calibration. This paper shows how the introduction of digital calibration for the ADC has a serious impact in the proposed test as calibration completely decorrelates signatures from the target specification in the presence of local mismatch.
