Detecting and correcting substrate process drift using machine learning

A method for training a machine learning model to predict a modification for a particular process recipe for a current substrate being processed at a manufacturing system, the method comprising:

• generating first training data for the machine learning model, wherein the first training data comprise historical data associated with a first set of prior substrates previously processed at the manufacturing system according to a first process recipe and a first set of historical metrology measurement values associated with each of the set of prior substrates;
• generating second training data for the machine learning model, wherein the second training data comprise historical data associated with a second set of prior substrates previously processed at the manufacturing system according to a second process recipe and a second set of historical metrology measurement values associated with each of the set of prior substrates;
• generating third training data for the machine learning model, wherein the third training data comprises an indication of a difference between the first process recipe and the second process recipe;
• and
• providing the first training data, the second training data, and the third training data to train the machine learning model to predict, for the particular process recipe for the current substrate being processed at the manufacturing system, which modification to the process recipe is to satisfy a drift criterion for a subsequent set of substrates that is to be processed after the current substrate.

A non-transitory computer readable storage medium comprising instructions that, when executed by a processing device, cause the processing device to:

• provide, as input to a trained machine learning model, data associated with processing each of a first set of substrates at a manufacturing system according to a process recipe;
• obtain one or more outputs from the trained machine learning model;
• determine, from the one or more outputs:
• an amount of drift of a first set of metrology measurement values for the first set of substrates from a target metrology measurement value, and
• process recipe modification data identifying one or more modifications to the process recipe and, for each of the modifications, an indication of a level of confidence that a respective modification to the process recipe satisfies a drift criterion for a second set of substrates;
• and
• responsive to identifying the respective modification with a level of confidence that satisfies a level of confidence criterion, update the process recipe based on the respective modification.