| CPC G06N 20/20 (2019.01) [G06F 11/0751 (2013.01); G06F 11/0766 (2013.01); G06N 3/06 (2013.01); G06N 5/04 (2013.01); G06N 20/00 (2019.01)] | 10 Claims |

|
1. A method comprising:
applying one or more patterns to the circuit;
identifying one or more failing patterns of a circuit as patterns that cause failing outputs from the circuit;
tracing failing outputs of the circuit to one or more candidate defect locations in the circuit;
classifying each failing pattern as a type-1 pattern, indicating that the failing pattern is explained by a single fault, a type-2 pattern, wherein the failing pattern is explained by multiple faults, or a type-3 pattern, wherein the failing pattern is not explained, indicating interacting defects are present in the circuit, the classifying comprising:
identifying one or more covers comprising a set of the candidate defect locations in the circuit that collectively explain all failing patterns;
assigning a fault type to each candidate defect location, the fault type selected from a group consisting of STUCK, CELL, BRIDGE and OPEN;
extracting, for a supervised machine learning model, a set of features from each candidate defect location based on patterns predicted by stuck-at and X-fault simulations of the candidate defect locations compared to observed failing patterns;
for each type-3 pattern:
applying a defect-level, fault type-specific machine learning model for each fault type to each candidate defect location, each fault type-specific machine learning model taking as input the features extracted from the candidate defect location and outputting a probability that the candidate defect location is the fault type which the fault type-specific machine learning model has been trained to detect; and
applying a cover-level machine learning model when the probability produced by at least one of the fault type-specific machine learning models returns a probability exceeding a predetermined threshold for at least one of the candidate defect locations in the cover, to classify the set of candidate defects in the cover as being correct or incorrect;
wherein a correct output from the cover-level machine learning model indicates that the cover represents actual defects in the circuit.
|