DIE LEVEL PRODUCT MODELING WITHOUT DIE LEVEL INPUT DATA

Abstract

A machine learning model for each die for imputing process control parameters at the die. The model is based on wafer sort parametric measurements at multiple test sites across the entire wafer, as well as yield results for the wafer. This allows for a better analysis of outlier spatial patterns leading to improved yield results.

Description

TECHNICAL FIELD

This application relates to modeling of semiconductor wafers, and more particularly, to modeling die level parameters.

BACKGROUND

It is very difficult to predict die level yield of semiconductor products for a variety of reasons. In previous generations, product yield was often defect limited and modeling die level yield required extensive optical defect data. Most efforts to model spatial patterns of yield or wafer sort parametric measurements with process control monitor (PCM) measurements focus on interpolating or extrapolating the PCM measurements to try and predict the PCM measurements at each wafer. To do this, electrical measurements of test structures on several wafers are taken as representative for a much larger group of wafers and then used to create a fixed parameter map. These efforts have generally been unsuccessful, since they could not account for the changing spatial patterns that are frequently responsible for bad wafers.

Further, due primarily to cost limitations, only a small percentage of wafers can be measured at each layer making die level prediction virtually impossible for most wafers. The yield of more recent generations is often driven more by parametric variation than defects, providing the opportunity to model yield directly from parametric measurements that can be performed on each die.

However, once again, cost provides an obstacle. Initial parametric measurements of scribe line structures, frequently known as either PCM or WAT (Wafer Acceptance Test) data, are normally performed at a limited number of sites, often only 5-10 sites, on each wafer. Predicting the wafer sort yield or wafer sort parametric on each die using PCM data is valuable, especially if the importance of each variable on each wafer can be specified. Since PCM measurements are well known to be correlated generally to parametric yield loss and wafer sort parametrics, many wafer level models have been successfully created using them to predict both wafer yield and mean value of parameters for each wafer. However, as noted, predicting yield at the die level is particularly difficult since the PCM measurements are normally available for only a few sites on the wafer. To predict die level yield, the most common approach is to impute the value of the PCM parameters at each die with interpolation or by fitting an assumed model form to the available PCM data. This has only been marginally successful due to the limited sites available and the uncertainty of the model forms for each PCM parameter.

Therefore, it would be desirable to predict the die yield or die level wafer sort parametric directly from the PCM parameters without the error associated with explicitly imputing the PCM parameter for each die.

BRIEF SUMMARY

A machine learning model is created for each die of a semiconductor wafer. Testing data is obtained from a plurality of testing sites distributed across the wafer. A die level map of the wafer is obtained. Based on the testing data and the die level map, each model is configured predict yield for its respective die, and to impute the process control parameters for the die.

DESCRIPTION OF DRAWINGS

FIG. 1 is a graphical plot illustrating a wafer-level correlation between predicted yield and actual yield for a plurality of wafer samples.

FIG. 2 is a graphical plot illustrating a lot-level correlation between predicted yield and actual yield for the plurality of wafer samples of FIG. 1.

FIG. 3 is a series of graphical plots illustrating die-level correlations between predicted yield and actual yield for a plurality of wafer lots.

FIG. 4 is a series of heat maps illustrating predicted spatial patterns for the wafer lots shown in FIG. 3.

FIG. 5 is a series of heat maps illustrating actual spatial patterns for the wafer lots shown in FIG. 3 and predicted in FIG. 4.

FIG. 6 is a flow chart of a general process for determining die-level values for process control parameters.

DETAILED DESCRIPTION

A wafer is a thin slice of semiconductor material used for the fabrication of integrated circuits. Many identical circuits are commonly formed on the wafer on individual dies, and when fabrication is complete, the dies are sliced apart at scribe lines to separate the individual circuits for packaging. Any structures close to the scribe lines will be destroyed during this process, making that area useless for creating product circuitry. However, these scribe line spaces are ideal for creating test structures that will be measured before slicing the dies.

Prior to slicing the dies, a testing protocol, such as wafer acceptance testing (WAT) or process control monitoring (PCM), is performed on scribe line test structures formed at various test sites on the wafer. In one example, nine test sites are measured on a wafer. For understanding the variation across the wafer, it would be ideal to locate some of the test sites near the edge of the wafer. However, it is usually preferred to locate the test sites away from wafer edges to improve the quality of the measurements.

To improve die-level yield predictions, a separate processor-based model can be built for each die on the wafer using the testing data obtained from each test site as a separate input variable to the model. Further, models could be built on a per site basis, a per die basis, a per wafer basis and/or a per lot basis. Modern machine learning technologies can be used to configure algorithmic-based software models that learn the complex non-linear relationships, initially from training sets of data, and updated from newly acquired data, to continuously learn about the relationships among the parameters to yield performance. For example, a neural network is an example of an implementation of a machine learning model, and XGBoost is another machine learning model based on extremely complex tree models.

The complex data relationships among process parameters and variables can be univariate or multivariate or both. By providing an analytical framework from which to better evaluate the various parameters and variables as representative of die level performance, consistently better yield performance overall can result. From the complex data relationships, the relevant input parameters for each die may be implicitly imputed by the model specific to that die.

The processor-based models could be desktop-based, i.e., standalone, or part of a networked system; but given the heavy loads of information to be processed and displayed with some interactivity, processor capabilities (CPU, RAM, etc.) should be current state-of-the-art to maximize effectiveness. In the semiconductor foundry environment, the Exensio® analytics platform is a useful choice for building GUI templates. In one embodiment, coding of the processing routines may be done using Spotfire® analytics software version 7.11 or above, which is compatible with Python object-oriented programming language, used primarily for coding machine language models.

The key is to learn and understand the complex non-linear relationships required to implicitly impute each PCM parameter for each die. This imputation is driven by the yield of the die so that the imputed PCM value may be implicitly predicted, since there is insufficient PCM data to explicitly predict each PCM value. Regardless of whether an implicit or explicit imputation form is chosen for the model, the accuracy is clearly demonstrated by the correlation between predicted and actual yield at the lot, wafer, and die levels, as shown in the figures.

FIG. 1 is a wafer level scatter plot 100 of predicted yield versus actual yield for a large number of wafers across multiple lots, with the linear correlation between predicted and actual yield shown as line 120. The linear correlation of predicted and actual yield has a cross-validated r-squared value of 0.561, indicating a correlation on average for all wafers of about 56% of predicted yield to actual yield.

A number of individual wafers labeled 101-109 are specifically referenced here and in the figures for illustrative purposes. For example, according to the wafer-level correlation shown in FIG. 1, wafer 103 has a relatively low correlation between predicted and actual yield, while wafers 102 and 109 have relatively high correlation between predicted and actual yield.

Referring to FIG. 2, a lot-level scatter plot 200 of the same wafers helps to drill down and provide additional input to the die-level model, illustrating additional variances at the lot level.

Since yield is a binary variable with a value of 0 for a failing die and a value of 1 for a passing die and the prediction is a number between 0 and 1 indicating the probability that the die will be yielding, it is more useful to compare the predicted yield with the average actual yield of each die and the die immediately adjacent to it. This is referred to as smoothing and the actual yield and predicted yield in all figures have been smoothed using this method.

The die-level correlations of predicted smoothed yield versus actual smoothed yield are shown in the series of plots in FIG. 3, which correspond to the 9 sample wafers 101-109. For example, the die correlation for wafer 109 has an average r-squared value of 0.757, indicating approximately 76% correlation between predicted and actual yield for the dies on this wafer; the die correlation for wafer 108 has an average r-squared value of 0.580, indicating approximately 58% correlation between predicted and actual yield for the dies on this wafer; the die correlation for wafer 107 has an average r-squared value of 0.057, indicating approximately 6% correlation between predicted and actual yield for the dies on this wafer; the die correlation for wafer 106 has an average r-squared value of 0.714, indicating approximately 71% correlation between predicted and actual yield for the dies on this wafer; the die correlation for wafer 105 has an average r-squared value of 0.293, indicating approximately 29% correlation between predicted and actual yield for the dies on this wafer; the die correlation for wafer 104 has an average r-squared value of 0.755, indicating approximately 76% correlation between predicted and actual yield for the dies on this wafer; the die correlation for wafer 103 has an average r-squared value of 0.430, indicating approximately 43% correlation between predicted and actual yield for the dies on this wafer; the die correlation for wafer 102 has an average r-squared value of 0.495, indicating approximately 50% correlation between predicted and actual yield for the dies on this wafer; and finally, the die correlation for wafer 101 has an average r-squared value of 0.331, indicating approximately 33% correlation between predicted and actual yield for the dies on this wafer.

These conclusions can be further confirmed by looking at the spatial correlations for the die models, as shown for wafers 101-109 in the predicted heat maps of FIG. 4 and the actual heat maps of FIG. 5. Even where yield results are lower, the models predict the final maps fairly well. The visual similarity of the wafer maps for wafers 103, 104, 105, 106, 108, and 109 are obvious. Wafers 101 and 107 are very different visually and the correlation plots in FIG. 3 show that these wafers have low values for spatial r-squared.

The concept is really a mapping of inputs to outputs, but based on the analysis of the complex relationship of the inputs to the desired outputs, in the primary case, yield. FIG. 6 is a flow chart illustrating a simple method for making die-level predictions of yield. In step 202, input data is acquired from each of the test sites formed across the wafer. In step 204, a die level map of the wafer is obtained, providing graphical details of features and locations. In step 206, for each die on the wafer, a machine learning model configured for the individual die predicts yield for that die based on the testing input data from all of the test sites. The implicit imputation of features is primarily driven by the yield performance, and the discovery of the inter-relationships between the input data from the entirety of the test sites on a single wafer and individual dies, learned and continuously updated by the machine learning model, establishes the correspondence to PCM values for individual die features. In optional step 208, the model can help to identify the relative importance of particular PCM variables to yield. Finally, in step 210, the models are deployed for use in evaluating spatial patterns for yield. Using the product yield and wafer sort parametric measurements as the intended response for the model allows PCM parameters for an individual die to be implicitly imputed, in particular, for modeling outlier spatial patterns.

In addition to imputing PCM values at multiple sites, this same technique can be used to incorporate other site level data, such as metrology and wafer level data such as tool, chamber, hold time, failure indicators, etc.

Thus, machine learning models can be configured to impute, for each die, the expected process control parameters based on wafer sort parametric measurements at multiple test sites across the entire wafer, as well as yield results for that wafer. This allows for a better analysis of outlier spatial patterns leading to improved yield results.

Claims

1. A method, comprising: obtaining a plurality of testing data from each of a respective plurality of test sites, the plurality of test sites distributed across a semiconductor wafer.obtaining a die level map of the wafer;based on the plurality of testing data and the die level map, for each of a plurality of dies on the wafer indicated on the die level map, predict yield for each of the plurality of dies using a respective machine learning model configured to model a respective one of the plurality of dies.