Trace-Based Transfer Learning for Substrate Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing manufacturing systems require multiple machine-learning models for different contexts, leading to scalability and reliability issues, as well as increased maintenance efforts due to the need for retraining and tuning models for each context.
Innovation Solution
Implementing trace-based transfer learning, where a transfer model generated from historical data of a source domain is used to modify trace data from a target domain, allowing a single machine-learning model to generate predictive data for multiple contexts without the need for retraining.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple machine-learning models are used for different contexts, then predictive data accuracy for each context is improved, but device complexity and maintenance effort increase
Solution Approach 1:
The patent implements a universal machine-learning model that can process trace data from multiple different contexts (substrate processing domains) by using domain identification and context-specific parameter selection. Instead of maintaining separate models for each context, a single model is trained to recognize domains and apply appropriate predictive parameters, thereby reducing model management complexity while maintaining predictive accuracy across diverse contexts.
2Reliability
If multiple machine-learning models are trained for different contexts, then reliability for each specific context is improved, but loss of time and computational resources increase
Solution Approach 1:
The patent performs preliminary domain identification and parameter selection before the actual predictive modeling occurs. By pre-training a universal model to recognize different substrate processing domains and pre-defining context-specific parameters, the system avoids the need for time-consuming retraining when new contexts are encountered. The domain identification mechanism is established in advance, enabling rapid adaptation to new contexts without extensive retraining.
3Adaptability or versatility
If multiple machine-learning models are maintained for different contexts, then adaptability to various contexts is improved, but ease of operation decreases
Solution Approach 1:
The patent introduces domain identification as an intermediary mechanism between the trace data input and the predictive model processing. This intermediary automatically determines the substrate processing domain and selects appropriate context-specific parameters, thereby simplifying the operation for users. Instead of requiring operators to manually configure different models for different contexts, the domain identification intermediary handles this automatically, maintaining high adaptability while improving ease of operation.
Data Source
AI summary
An electronic device manufacturing system configured identify a machine-learning model trained to generate analytic or predictive data for a first substrate processing domain associated with a type of substrate processing system. The system is further configured to obtain first trace data pertaining to the first domain used to train the machine-learning model. The system is further configured to a transfer model for a second substrate processing domain associated with the type of substrate processing system. The transfer model is generated based on the first trace data pertaining to the first substrate processing domain and second trace data pertaining to the second substrate processing domain. Using the transfer model, at least one of the machine-learning model or current trace data associated with the second substrate processing domain is modified to enable the machine-learning model to generate analytic or predictive data associated with the second substrate processing domain.