Method and apparatus for self-learning and self-improving a semiconductor manufacturing tool

• US 8,396,582 B2
• Filed: 01/29/2010
• Issued: 03/12/2013
• Est. Priority Date: 03/08/2008
• Status: Active Grant

• Current Assignee: Tokyo Electron Limited
• Original Assignee: Tokyo Electron Limited
• Inventors: Kaushal, Sanjeev, Patel, Sukesh Janubhai, Sugishima, Kenji
• Primary Examiner(s): KASENGE, CHARLES R
• Application Number:
  US12/697,121
• Publication Number:
  US 20100138026A1
• Time in Patent Office: 1,138 Days
• Field of Search: 700/28, 700/47, 700/52, 700/100, 700/103, 700/105, 700/108, 700/109, 700/110, 700/121, 702/81, 702/82, 702/83, 702/84, 702/182, 702/183, 702/184, 438/5, 438/14, 438/16, 438/17, 706/12, 706/13, 706/16, 706/23, 706/25
• US Class Current: 700/104
• CPC Class Codes: G05B 13/0265 the criterion being a learning criterion
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G06N 5/04 Inference or reasoning models

Abstract

System(s) and method(s) for optimizing performance of a manufacturing tool are provided. Optimization relies on recipe drifting and generation of knowledge that capture relationships among product output metrics and input material measurement(s) and recipe parameters. Optimized recipe parameters are extracted from a basis of learned functions that predict output metrics for a current state of the manufacturing tool and measurements of input material(s). Drifting and learning are related and lead to dynamic optimization of tool performance, which enables optimized output from the manufacturing tool as the operation conditions of the tool changes. Features of recipe drifting and associated learning can be autonomously or externally configured through suitable user interfaces, which also can be drifted to optimize end-user interaction.

Claims

1. An autonomous biologically based learning system, comprising:

a manufacturing tool that produces an asset;

a drift component that modifies a manufacturing recipe processed by the manufacturing tool at least in part using a set of driving variables and a predetermined probability distribution function to generate one or more adjusted manufacturing recipes for the asset, wherein the set of driving variables determine a particular sequence to modify a set of recipe parameters associated with the manufacturing recipe;

an objective autonomous learning engine that infers one or more functions for the manufacturing tool based on the modified manufacturing recipe processed by the manufacturing tool, wherein the one or more functions predict asset output metrics for the produced asset; and

an autonomous optimization engine that extracts a set of updated recipe parameters from a set of input measurements and the one or more inferred functions to generate an adjusted recipe within a predefined tolerance of a target value for the asset output metrics.

2. The system of claim 1, wherein the asset includes one or more of semiconductor devices, the semiconductor devices comprising semiconductor substrates possessing integrated circuits (ICs), plasma-discharge based flat panel displays (FPDs), organic light emitting diode (OLED)-based FPDs, liquid crystal display (LCD)-based FPDs, or elements thereof which comprise one or more of thin-film-transistor active matrix arrays, color filters, or polarizers.

3. The system of claim 1, wherein the drift component randomly modifies the manufacturing recipe at least in part using the set of driving variables and the predetermined probability distribution function.

4. The system of claim 1, wherein the drift component includes a formatter component that generates a recipe genealogy that includes at least one of the one or more adjusted manufacturing recipes, input material data linked to respective adjusted manufacturing recipes, output data related to respective adjusted manufacturing recipes, or a set of operational conditions for the manufacturing tool associated with respective adjusted manufacturing recipes.

5. The system of claim 4, wherein the set of operational conditions for the manufacturing tool associated with respective adjusted manufacturing recipes includes at least one of elapsed active duty, preventive maintenance (PM) records, or repair records.

6. The system of claim 1, wherein, to infer the one or more functions, the objective autonomous learning engine relaxes one or more constraints for the asset output metrics.

7. The system of claim 6, wherein the relaxation of the one or more constraints is dictated at least in part by a ranking of weights for the asset output metrics.

8. The system of claim 1, wherein the autonomous optimization engine solves an inverse problem to extract the set of recipe parameters, the inverse problem solved through at least one of genetic algorithm, Monte Carlo simulation, simulated annealing, or Tabu search.

9. The system of claim 1, wherein the autonomous optimization engine configures repair procedures in response to a determination that the adjusted recipe is not generated within the predefined tolerance of the target value for the asset output metrics.

10. The system of claim 1, further comprising a component that renders a user interface to configure at least one of the set of driving variables.

11. The system of claim 6, further comprising a component that delivers the one or more functions to at least one disparate manufacturing tool to learn at least one function in the disparate manufacturing tool.

12. A method, comprising:

employing at least one processor to execute code instructions retained in a memory, the code instructions when executed by at least the one processor carry out acts, comprising;

determining a particular sequence for modifying one or more parameters associated with an initial recipe;

modifying the initial recipe at least in part using the particular sequence and a predetermined probability distribution function to fabricate a product that fulfills target product output;

learning a set of relationships associated with product output through the modifying of the initial recipe;

generating an adjusted recipe to accomplish the target product output based at least in part on the set of learned relationships;

evaluating whether the target product output is fulfilled in response to determining that generation of the adjusted recipe is successful; and

continuing the modifying in response to determining that the evaluation yields an unsuccessful outcome.

13. The method of claim 12, further comprising effecting a maintenance procedure in response to unsuccessfully generating the adjusted recipe.

14. The method of claim 12, wherein the modifying includes:

setting the initial recipe as a current recipe;

randomizing the one or more parameters in the current recipe;

processing via a manufacturing tool an adjusted recipe obtained in response to the randomizing;

adding the adjusted recipe to a genealogy of recipes created through randomizing, and setting the adjusted recipe as a current recipe in response to determining that the adjusted recipe is to be randomized; and

supplying the genealogy of recipes created through the randomizing, and data associated with the genealogy in response to determining that the adjusted recipe is not to be randomized.

15. The method of claim 12, wherein the generating includes:

receiving the set of learned relationships that predict product output for a set of recipes;

receiving input material measurements;

extracting recipe parameters for the adjusted recipe based at least in part on a set of learned functions or the received input material measurements;

supplying the adjusted recipe at least in part with the extracted recipe parameters in response to determining that the predicted product output for each learned relationship in the set of learned relationships satisfies output constraints.

16. The method of claim 15, in response to determining that the predicted product output for each learned relationship in the set of learned relationships does not satisfy output constraints, further comprising:

relaxing a set of output constraints; and

extracting recipe parameters for the adjusted recipe based at least in part on the received set of learned functions or the received input material measurements in response to determining that a current number of relaxed constraints is below a predetermined threshold.

17. The method of claim 16, further comprising configuring a maintenance repair procedure in response to determining that the current number of relaxed constraints is below the predetermined threshold.

18. The method of claim 14, wherein the learning includes:

receiving production data generated through execution of one or more modified recipes;

learning a functional relationship among a product output metric, material input measurements, and recipe parameters of the manufacturing tool based on the received production data; and

accepting the learning functional relationship in response to determining that predictions of the product output are within a specified quality.

19. The method of claim 15, wherein the extracting includes solving an inverse problem through at least one of genetic algorithm, Monte Carlo simulation, simulated annealing, or Tabu search.

20. The method of claim 18, wherein the accepting includes at least one of delivering the learned functional relationship or storing the learned functional relationship.

21. A non-transitory computer-readable storage medium comprising a set of code instruction retained therein that, in response to execution, cause a computing system including at least one processor to perform operations, comprising:

determining a particular sequence for modifying one or more parameters associated with an initial recipe;

modifying the initial recipe at least in part using the particular sequence and a predetermined probability distribution function to fabricate a product that fulfills target product output;

learning a set of relationships associated with product output through the modifying of the initial recipe;

generating an adjusted recipe to accomplish the target product output based at least in part on the set of learned relationships;

evaluating whether the target product output is fulfilled in response to determining that generation of the adjusted recipe is successful; and

continuing the modifying in response to determining that the evaluation results in an unsuccessful outcome.