Selective inclusion/exclusion of semiconductor chips in accelerated failure tests

1. A method, comprising:

obtaining and validating historical input data from wafer acceptance testing, circuit probe testing, and final testing for one or more previous production runs for making a plurality of semiconductor chips;

detecting one or more anomalies in the historical input data;

identifying one or more features of interest based on the detection of anomalies;

training at least one model using the historical input data as a training dataset, the model configured to determine from a set of input data whether any of the plurality of semiconductor chips have detected anomalies in the features of interest and should be subjected to further testing;

creating at least one relabeled model from the at least one model:

determining a first set of data in the training dataset that are adjacent or similar to a second set of data, wherein the second set of data identifies chips having a first feature of interest that is determined to be defective and the first set of data identifies chips wherein the first feature of interest that is determined to be marginal but acceptable; and

removing the first set of data from the training dataset to form a relabeled dataset;

training a final model to replace the at least one model using the relabeled dataset;

deploying the final model to evaluate current input data from wafer acceptance testing, circuit probe testing, and final testing for a current production run; and

predicting, by the final model using the current input data, for every chip in the current production run, whether the chip will exhibit an anomaly in a feature of interest that exceeds a defined threshold for acceptability in the feature of interest, and therefore should be subject to further testing.

2. The method of claim 1 , further comprising:

detecting one or more anomalies in the historical input data using any of a plurality of outlier detection methods.

3. The method of claim 2 , further comprising:

detecting one or more anomalies in the historical input data using any of a plurality of statistical methods to identify one or more atypical data points in the historical input data.

4. The method of claim 2 , further comprising:

grouping together selected features of interest that exhibit detected anomalies;

calculating aggregated statistics for the grouped features;

creating one or more new features of interest based on the aggregated statistics; and

retraining the at least one model to consider the new features of interest.

5. The method of claim 1 , further comprising:

modifying the training dataset by substituting replacement data having non-anomalous values for anomalous data.

6. The method of claim 5 , wherein the replacement data is an upper acceptable value, a lower acceptable value, or a median value, for the feature of interest.

7. The method of claim 1 , further comprising:

creating a plurality of the relabeled models from the at least one model;

removing the first set of data from the training dataset to form the relabeled dataset only when a predefined number of the relabeled models make the same determination.