Die level product modeling without die level input data

1. A method, comprising:

obtaining testing data from each of a respective plurality of test sites formed and distributed across a semiconductor wafer, prior to slicing a multiplicity of dies from the wafer;

obtaining a die level map of the semiconductor wafer that includes graphical details regarding a multiplicity of semiconductor features and corresponding locations for each feature on the semiconductor wafer; and

for each of the multiplicity of dies on the semiconductor wafer:

providing the testing data and the die level map as input data to a multiplicity of neural networks, each neural network corresponding to a respective one of the multiplicity of dies and each neural network programmed with instructions to (i) determine non-linear relationships among the input data for the respective die, (ii) impute from the determined non-linear relationship among the input data a plurality of input process parameters relevant to the respective die, and (iii) predict yield for the respective die from the imputed input process parameters, wherein each neural network is initially configured from training sets of input data.

2. The method of claim 1 , further comprising:

identifying by the neural network a relative importance of each of the multiplicity of process parameters to the yield prediction for the respective die.

3. A method, comprising:

obtaining testing data for a plurality of process control parameters from a plurality of test sites formed on and distributed across a semiconductor wafer, the testing data obtained prior to slicing a multiplicity of dies from the wafer;

obtaining a die level map of the semiconductor wafer, the map including graphical details regarding a multiplicity of semiconductor features formed on each die and corresponding locations on the wafer for each of the multiplicity of features;

providing the testing data and the die level map as input data to a multiplicity of neural networks, each one of the multiplicity of neural networks corresponding to a respective one of the multiplicity of dies,

wherein each of the multiplicity of neural networks is programmed with instructions to learn and identify non-linear relationships between each of the plurality of process control parameters and yield performance on the basis of the testing data and the die level map, and to impute each of the plurality of process control parameters for each respective die on the basis of the determined non-linear relationships, each of the multiplicity of neural networks is initially configured from training sets of input data; and

deploying the method into a semiconductor processing environment.

4. The method of claim 3 , further comprising:

each neural network is configured to determine the correspondence between individual ones of the process control parameters and individual ones of the semiconductor features.

5. The method of claim 3 , wherein each neural network is initially configured on the basis of training sets of input data.

6. The method of claim 3 , wherein each neural network is continuously updated from newly acquired sets of input data from the semiconductor processing environment to continuously learn and identify the non-linear relationships.

7. The method of claim 3 , further comprising evaluating spatial patterns on die level maps for correspondence to yield performance.

8. The method of claim 7 , further comprising modeling outlier spatial patterns based on the imputed process control parameters.

9. The method of claim 3 , further comprising:

forming at least some of the plurality of test sites on a plurality of scribe lines located between the dies on the wafer for indicating where to slice the dies.

10. The method of claim 3 , further comprising:

forming at least some of the plurality of test sites near edges of the wafer.

11. A non-transitory computer-readable medium having instructions which, when executed by a processor, cause the processor to:

obtain testing data for a plurality of process control parameters from a plurality of test sites formed on and distributed across a semiconductor wafer, the testing data obtained prior to slicing a multiplicity of dies from the wafer;

obtain a die level map of the semiconductor wafer, the map including graphical details regarding a multiplicity of semiconductor features formed on each die and corresponding locations on the wafer for each of the multiplicity of features; and

provide the testing data and the die level map as input data to a multiplicity of neural networks, each one of the multiplicity of neural networks corresponding to a respective one of the multiplicity of dies,

wherein each of the multiplicity of neural networks is programmed with instructions to learn and identify non-linear relationships between each of the plurality of process control parameters and yield performance on the basis of the testing data and the die level map, and to impute each of the plurality of process control parameters for each respective die on the basis of the determined non-linear relationships, each of the multiplicity of neural networks is initially configured from training sets of input data.

12. A method, comprising:

obtaining testing data for a plurality of process control parameters from a plurality of test sites formed on and distributed across a semiconductor wafer, the testing data obtained prior to slicing a multiplicity of dies from the wafer;

obtaining a die level map of the semiconductor wafer, the map including graphical details regarding a multiplicity of semiconductor features formed on each die and corresponding locations on the wafer for each of the multiplicity of features;

providing the testing data and the die level map as input data to a multiplicity of machine learning tree models, each one of the multiplicity of machine learning tree models corresponding to a respective one of the multiplicity of dies; and

wherein each of the multiplicity of machine learning tree models is programmed with instructions to learn non-linear relationships between each of the plurality of process control parameters and yield performance on the basis of the testing data and the die level map, and to impute each of the plurality of process control parameters for each respective die on the basis of the determined non-linear relationships, each of the multiplicity of machine learning tree models is initially configured from training sets of input data.