Neural Network Control Relearning for Lithography Accuracy Drift
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Control accuracy deteriorates over time due to changes in the state of a controlled object, even when a neural network is optimized, as the state of the object changes, affecting the performance of control apparatuses.
Innovation Solution
A management apparatus that uses reinforcement learning to redecide parameter values of a neural network controller when the reward from control results does not meet a predetermined criterion, by iteratively adjusting and optimizing the parameter values to maintain control accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a neural network is optimized at a given time, then control accuracy is improved, but control accuracy deteriorates over time as the state of the controlled object changes
Solution Approach 1:
The patent implements dynamic relearning of neural network parameters based on changes in the controlled object's state. The system continuously monitors state changes and triggers relearning when necessary, making the control system adaptive rather than static. This resolves the contradiction by allowing the neural network to maintain optimal performance despite temporal changes in the controlled object.
Solution Approach 2:
The patent employs feedback mechanisms where the control results are evaluated and used to determine whether relearning should be performed. The system calculates changes in the controlled object's state and uses this feedback to decide on parameter reoptimization, ensuring control accuracy is maintained throughout operation.
2Productivity
If reinforcement learning is used to decide parameter values, then control performance is improved, but additional computational complexity and time are required for relearning
Solution Approach 1:
The patent performs preliminary evaluation of state changes before initiating full relearning. By calculating the magnitude of state changes and comparing them against thresholds, the system determines whether relearning is necessary, avoiding unnecessary computational overhead while maintaining control performance when needed.
Solution Approach 2:
The patent selectively updates neural network parameters based on the detected state changes. Rather than always performing complete relearning, the system adjusts parameters only when state changes exceed predetermined thresholds, reducing computational time while maintaining control effectiveness.
Data Source
AI summary
A management apparatus includes a learning device. The learning device is configured to, in a case where a reward obtained from a control result of a controlled object by a controller configured to control the controlled object using a neural network, for which a parameter value is decided by reinforcement learning, does not satisfy a predetermined criterion, redecide the parameter value by reinforcement learning.