Method and apparatus for predicting yield of semiconductor devices

Patent metadata

Field	Value
Publication number	US-12198061-B2
Application number	US-202117351975-A
Country	US
Kind code	B2
Filing date	Jun 18, 2021
Priority date	Nov 27, 2020
Publication date	Jan 14, 2025
Grant date	Jan 14, 2025

Abstract

A method for predicting the yield of manufacturing semiconductor devices includes steps of: acquiring defect data of semiconductor devices to be predicted, wherein the semiconductor devices to be predicted include finished semiconductor devices and semi-finished semiconductor devices, and the defect data indicates a defect type and location of at least one defect of the semiconductor devices; inputting the defect data into a pre-trained yield prediction model, wherein the yield prediction model includes a neural network structure and a classification structure, the neural network structure is used to extract defect feature vectors from the defect data, and the classification structure is used to output classification results of qualified or unqualified yield according to the defect feature vectors; and determining, by the yield prediction model, classification results of qualified or unqualified yield of the semiconductor devices.

First claim

What is claimed is:

1. A method for predicting yield of manufacturing semiconductor devices, comprising:
acquiring defect data of the semiconductor devices, wherein the semiconductor devices comprise finished semiconductor devices and semi-finished semiconductor devices, and wherein the defect data indicates a defect type and a location of at least one defect of the semiconductor devices;

inputting the defect data into a pre-trained yield prediction model, wherein the yield prediction model comprises a neural network structure and a classification structure, wherein the neural network structure extracts defect feature vectors from the defect data, wherein the defect feature vectors define qualified or unqualified yield, and wherein the classification structure sort out classification results of the qualified or the unqualified yield; and
outputting, by the yield prediction model, the classification results of qualified or unqualified yield of the semiconductor devices.

2. The yield predicting method of claim 1 , further comprising steps of:
providing a plurality of semiconductor device test samples;
acquiring sample defect data and sample yield data of the plurality of semiconductor device test samples, wherein the sample defect data comprises a defect type and location of at least one defect of the plurality of semiconductor device test samples, and the sample yield data indicates whether a yield of the plurality of semiconductor device test samples is qualified; and
training the yield prediction model with the sample defect data as input information and the sample yield data as output information.

3. The yield predicting method of claim 2 , wherein the training the yield prediction model comprises:
training each neural network layer of a pre-training structure without supervision, using the sample defect data of the plurality of semiconductor device test samples as the input information of a first neural network layer of the pre-training structure and output information of a last neural network layer of the pre-training structure, wherein each said neural network layer maps feature information to a next neural network layer as much as possible; and
removing at least the last neural network layer of the pre-training structure to obtain the neural network structure, wherein the neural network structure comprises a plurality of neural network layers.

4. The yield predicting method of claim 3 , wherein each one of the plurality of neural network layers comprises at least one mutually independent neuron, wherein each neuron is adapted to calculate corresponding output information according to its input information and synaptic weight and input the output information to each neuron of a next one of the plurality of neural network layers, wherein the output information is a continuous nonlinear function of the input information and the synaptic weight; and
wherein the training each neural network layer of the pre-training structure comprises:
training the synaptic weight of each neuron according to the input information and output information of the neuron of each one of the plurality of neural network layers to train the neural network.

5. The yield predicting method of claim 3 , wherein the pre-training structure comprises a plurality of front neural network layers and a plurality of back neural network layers, wherein dimensions of the plurality of front neural network layers are reduced layer by layer, dimensions of the plurality of back neural network layers are increased layer by layer, and wherein a step of obtaining the neural network structure comprises:
removing the plurality of back neural network layers of the pre-training structure, and using the plurality of front neural network layers as the neural network structure.

6. The yield predicting method of claim 2 , wherein the classification structure comprises a support vector machine (SVM) structure, wherein the SVM structure is adapted to determine a hyperplane that can divide all data of the defect feature vectors, so as to realize a classification of the defect feature vectors, wherein the hyperplane makes a distance between each of all the data of the defect feature vectors and the hyperplane a shortest.

7. The yield predicting method of claim 6 , wherein the SVM structure comprises an SVM network, wherein the SVM network is configured behind a last neural network layer of the neural network structure to form the yield prediction model, and wherein the training the yield prediction model further comprises:
reversely training the plurality of neural network layers of the neural network structure with supervision, using the defect feature vectors outputted by the last neural network layer of the neural network structure as input information of the SVM network and the sample yield data as output information of the SVM network, to finely adjust the yield prediction model.

8. The yield predicting method of claim 7 , wherein the training the yield prediction model further comprises:
determining abnormal points outputted by a finely adjusted yield prediction model according to a baseline fail bin count at a normal level of a long-term trend chart;
removing data of the abnormal points from the sample defect data, and performing null filling on data locations of the abnormal points with an average value of defects of multiple lots of adjacent samples; and
adjusting a penalty coefficient, a gamma value, and/or a weight of the support vector machine structure according to the accuracy, detection rate and error report rate of the abnormal points to optimize the yield prediction model.

9. The yield predicting method of claim 2 , wherein the acquiring sample defect data of the plurality of semiconductor device test samples comprises:
screening multiple lots of semiconductor device test samples on a production line according to a key layer to select multiple lots of typical semiconductor device test samples, wherein the key layer is determined by a location of at least one defect that has a high yield influencing weight, wherein each lot of the multiple lots of semiconductor device test samples comprises a plurality of wafers, wherein each of the plurality of wafers comprises a multi-layer structure;
screening the multiple lots of typical semiconductor device test samples to select a plurality of typical wafers;
performing defect analysis on the plurality of typical wafers to determine at least one defect of each typical wafer; and
counting a defect type and location of the at least one defect of each typical wafer, to determine the sample defect data of the multiple lots of typical semiconductor device test samples.

10. The yield predicting method of claim 9 , wherein the acquiring sample defect data of the plurality of semiconductor device test samples further comprises:
screening defects of the multiple lots of semiconductor device test samples according to yield influencing weights of defects of various defect types to determine types of killer defects;
screening the defects of the multiple lots of semiconductor device test samples according to the yield influencing weights of defects at respective locations to determine the key layer; and
screening the sample defect data of the multiple lots of semiconductor device test samples according to the types of killer defects and the key layer to determine optimized sample defect data.

11. The yield predicting method of claim 9 , further comprising:
in response to missing sample defect data of any semiconductor device sample, calculating an average value of the sample defect data of multiple adjacent layers during performing missing data optimization algorithm on the missing

Assignees

• Shanghai Huali Integrated Circuit Corp

Inventors

• Chen Shanshan
• Chen Hunglin
• Long Yin

Classifications

• H10P74/23

  characterised by multiple measurements, corrections, marking or sorting processes · CPC title

• G06N3/0455

  Auto-encoder networks; Encoder-decoder networks · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06N3/0895

  Weakly supervised learning, e.g. semi-supervised or self-supervised learning · CPC title

• G06N3/0499

  Feedforward networks · CPC title