# US 2021/0279388 A1 · Predicting Die Susceptible to Early Lifetime Failure

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2021/0279388 A1
- Document ID: US20210279388A1
- Status: Published application
- Kind: A1
- Application number: 17/189,621
- Filed: 2021-03-02
- Published: 2021-09-09
- Earliest claimed: 2020-03-03
- Assignees: PDF Solutions, Inc.
- Applicants: PDF Solutions, Inc.
- Inventors: Richard Burch; Qing Zhu; Keith Arnold
- Web page: https://patentgrep.com/patent/US20210279388A1

## CPC

- G06F 30/27: using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
- G06F 30/398: Design verification or optimisation, e.g. using design rule check [DRC], layout versus schematics [LVS] or finite element methods [FEM] (optical proximity correction [OPC] design processes G03F1/36)
- G06F 2119/02: Reliability analysis or reliability optimisation; Failure analysis, e.g. worst case scenario performance, failure mode and effects analysis [FMEA]
- G06F 2119/04: Ageing analysis or optimisation against ageing
- G06F 2119/22: Yield analysis or yield optimisation
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)

## Abstract

Semiconductor yield is modeled at the die level to predict die that are susceptible to early lifetime failure (ELF). A first die yield calculation is made from parametric data obtained from wafer testing in a semiconductor manufacturing process. A second die yield calculation is made from die location only. The difference between the first die yield calculation and the second die yield calculation is a prediction delta. Based on an evaluation of the first die yield calculation and the prediction delta, the likelihood of early lifetime failure can be identified and an acceptable level of die loss can be established to remove die from further processing.

## Claims

### Claim 1

A method, comprising:
predicting a first yield value for each of a first plurality of die in a first identified parameter group using a first machine learning model configured to predict the first yield value using data values obtained from parametric testing of a semiconductor wafer on which the first plurality of die are formed, each of the first plurality of die have data values for each of a plurality of parameters in the first identified parameter group;
assigning the first yield value to each of the first plurality of die;
predicting a second yield value for each of the first plurality of die in the first parameter group using a second machine learning model configured to predict the second yield value using a location of each of the first plurality of die;
determining a prediction delta by subtracting the predicted second yield value from the predicted first yield value;
establishing a threshold for acceptable die loss based on a relationship between the predicted first yield value and the prediction delta; and
removing a portion of the first plurality of die from further processing in accord with the threshold for acceptable die loss.

### Claim 2 (depends on 1)

The method of claim 1, further comprising:
predicting the first yield value for additional pluralities of die, each additional plurality of die organized into one of a plurality respective parameter groups, each of the additional pluralities of die having data values for each of a plurality of parameters in each of the respective parameter groups; and
predicting the second yield value for each of the additional pluralities of die in respective parameter groups using a location of each die.

### Claim 3 (depends on 1)

The method of claim 1, further comprising:
identifying a plurality of parameter groups, each identified parameter group including a respective plurality of die each having data values for each of a plurality of parameters in the identified parameter group;
predicting the first yield value for each of the respective plurality of die in each identified parameter group; and
predicting the second yield value for each of the respective plurality of die in each identified parameter group using a location of each die.

### Claim 4 (depends on 1)

The method of claim 1, further comprising:
estimating a maximum limit, respectively, for each of the plurality of parameters;
for each die, comparing each of the plurality of parameters to the respective maximum limit;
removing die having a parameter that exceeds the respective maximum limit.

### Claim 5 (depends on 1)

The method of claim 1, further comprising:
evaluating the relationship between the predicted first yield value and the prediction delta in order to establish the threshold value.

### Claim 6 (depends on 5)

The method of claim 5, further comprising:
sorting the predicted first yield values from lowest to highest;
calculating a yield percentile of the predicted first yield value associated with each of the first plurality of die;
sorting the prediction deltas from lowest to highest;
calculating a delta percentile of the prediction delta associated with the each of the first plurality of die; and
identifying a minimum percentile value as between the yield percentile and the delta percentile;
selecting the minimum percentile value as the threshold value.

### Claim 7 (depends on 5)

The method of claim 5, further comprising:
setting the threshold for acceptable die loss for a condition when the predicted first yield value is less than a defined limit and when the prediction delta is a negative value.

### Claim 8 (depends on 5)

The method of claim 5, further comprising:
setting the threshold for acceptable die loss for a condition when the predicted first yield value is less than a defined limit and when the prediction delta is a positive value.

### Claim 9 (depends on 1)

The method of claim 1, wherein the first machine learning model is further configured using data obtained from fabrication processing of the semiconductor wafer.

### Claim 10 (depends on 1)

The method of claim 1, wherein the first machine learning model is further configured using data obtained from packaging of the semiconductor wafer.
