Optimizing semiconductor manufacturing processes using machine learning
In some embodiments, a computer-implemented method of controlling a semiconductor manufacturing process is provided. A computing system generates predicted metrology values for a current run and a next run by providing metrology forecast inputs to a metrology forecast model. The computing system generates an updated recipe for executing at least one semiconductor manufacturing process step using the predicted metrology values for the current run and the next run.
1. A computer-implemented method of controlling a semiconductor manufacturing process, the method comprising:
generating, by a computing system, predicted metrology values for a current run and predicted metrology values for a next run by providing metrology forecast inputs associated with the current run to a metrology forecast model;
providing, by the computing system, at least the predicted metrology values for the current run and the next run to an actor model to generate an updated recipe for executing at least one semiconductor manufacturing process step; and
using the updated recipe to control at least one manufacturing device during the at least one semiconductor manufacturing process step.
2. The computer-implemented method of claim 1 , wherein the at least one semiconductor manufacturing process step includes at least one of thin film deposition, photolithography, etching, overlay correction, or chemical mechanical planarization.
3. The computer-implemented method of claim 1 , wherein providing the metrology forecast inputs associated with the current run to the metrology forecast model includes providing at least one of process input values, trace statistic values, exogenous values, apriori values, or measured metrology values;
wherein the process input values include values of control inputs used to control a process step;
wherein the trace statistic values include at least one of a scalar metric extracted from a sensor trace or a latent dimension extracted from the sensor trace by a deep neural network autoencoder;
wherein the exogenous values include at least one of a timestamp of a run started, an ambient temperature, or a relative humidity;
wherein the apriori values include at least one of a wafer number, a chamber accumulation counter value, a hot plate identifier, and a measurement value from a previous process step; and
wherein the measured metrology values include at least one of a thickness, a stress, a refractive index, or an etch critical dimension.
4. The computer-implemented method of claim 1 , further comprising retraining alpha parameters of the actor model after a predetermined period of time or a predetermined number of runs.
5. The computer-implemented method of claim 1 , wherein providing at least the predicted metrology values for the current run and the next run to the actor model to generate the updated recipe includes:
providing process model inputs to a process model to determine predicted process outputs; and
evaluating the predicted process outputs and the predicted metrology values for the current run and the next run using a cost function to determine the updated recipe.
6. The computer-implemented method of claim 5 , wherein the process model inputs include one or more of a deposition time value, a high frequency (HF) power value, an argon flow value, a pedestal gap value, a dosing value, an etch time value, or an etch gas flow value.
7. The computer-implemented method of claim 5 , wherein the process model is linearized about an operating point in a space of the process model inputs.
8. The computer-implemented method of claim 5 , wherein output of the process model includes a prediction for each output dimension within a space of the process model inputs.
9. The computer-implemented method of claim 5 , further comprising retraining the process model in response to determining that a variance in an independent input space exceeds a threshold variance proportional to known model parameter uncertainty.
10. The computer-implemented method of claim 1 , further comprising retraining the metrology forecast model in response to obtaining measured metrology values.
11. A non-transitory computer-readable medium having computer-executable instructions stored thereon that, in response to execution by one or more processors of a computing system, cause the computing system to perform actions for controlling a semiconductor manufacturing process, the actions comprising:
generating, by the computing system, predicted metrology values for a current run and predicted metrology values for a next run by providing metrology forecast inputs associated with the current run to a metrology forecast model;
providing, by the computing system, at least the predicted metrology values for the current run and the next run to an actor model to generate an updated recipe for executing at least one semiconductor manufacturing process step; and
using the updated recipe to control at least one manufacturing device during the at least one semiconductor manufacturing process step.
12. The non-transitory computer-readable medium of claim 11 , wherein providing the metrology forecast inputs associated with the current run to the metrology forecast model includes providing at least one of process input values, trace statistic values, exogenous values, apriori values, or measured metrology values;
wherein the process input values include values of control inputs used to control a process step;
wherein the trace statistic values include at least one of a scalar metric extracted from a sensor trace or a latent dimension extracted from the sensor trace by a deep neural network autoencoder;
wherein the exogenous values include at least one of a timestamp a run started, an ambient temperature, or a relative humidity;
wherein the apriori values include at least one of a wafer number, a chamber accumulation counter value, a hot plate identifier, and a measurement value from a previous process step; and
wherein the measured metrology values include at least one of a thickness, a stress, a refractive index, or an etch critical dimension.
13. The non-transitory computer-readable medium of claim 11 , wherein providing at least the predicted metrology values for the current run and the next run to the actor model to generate the updated recipe includes:
providing process model inputs to a process model to determine predicted process outputs; and
evaluating the predicted process outputs and the predicted metrology values for the current run and the next run using a cost function to determine the updated recipe.
14. The non-transitory computer-readable medium of claim 13 , wherein the process model inputs include one or more of a deposition time value, a high frequency (HF) power value, an argon flow value, a pedestal gap value, a dosing value, an etch time value, or an etch gas flow value.
15. The non-transitory computer-readable medium of claim 13 , wherein the process model is linearized about an operating point in a space of the process model inputs.
16. The non-transitory computer-readable medium of claim 13 , wherein output of the process model includes a prediction for each output dimension within a space of the process model inputs.
17. The non-transitory computer-readable medium of claim 13 , wherein the actions further comprise retraining the process model in response to determining that a variance in an independent input space exceeds a threshold variance proportional to known model parameter uncertainty.
18. The non-transitory computer-readable medium of claim 11 , wherein the at least one semiconductor manufacturing process step includes at least one of thin film deposition, photolithography, etching, overlay correction, or chemical mechanical planarization.
19. The non-transitory computer-readable medium of claim 11 , wherein the actions further comprise retraining alpha parameters of the actor model after a predetermined period of time or a predetermined number of runs.
20. The non-transitory computer-readable medium of claim 11 , wherein the actions further comprise further comprising retraining the metrology forecast model in response to obtaining measured metrology values.