Reinforcement Learning Model for Adaptive Process Recipe Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional semiconductor processing apparatuses struggle to dynamically adjust control values for process recipes based on the current shape of the processing object, leading to inefficiencies in achieving target processing shapes.
Innovation Solution
A model generation method using reinforcement learning that acquires variable groups, state data, and target data to evaluate and generate models for constructing process recipes, recommending variable groups based on the current shape of the processing object.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional process recipes with fixed control values are used, then the process can be executed with simple control, but the processing cannot be dynamically adjusted to achieve target shapes efficiently
Solution Approach 1:
The patent applies dynamics by transitioning from static, pre-defined process recipes to dynamic, adaptive process recipes. The reinforcement learning model enables the system to adapt control values in real-time based on the current shape of the processing object, allowing the process recipe to evolve dynamically during execution rather than following fixed predetermined parameters
Solution Approach 2:
The patent implements feedback mechanisms by using the current shape information of the processing object to adjust subsequent processing steps. The reinforcement learning model continuously receives feedback about the processing object's state and modifies control values accordingly, creating a closed-loop control system that adapts to actual processing conditions
2Loss of time
If reinforcement learning models are trained offline, then the model generation time is reduced, but the model may not adapt to real-time variations in processing objects
Solution Approach 1:
The patent applies preliminary action by training the reinforcement learning model offline before actual processing begins. This pre-training phase allows the model to learn from extensive simulation data and establish baseline knowledge, so that during real-time processing, the model can quickly adapt to specific cases without requiring extensive computation, thus reducing online processing time while maintaining reliability
3Measurement precision
If extensive simulation data is used for training, then the model accuracy improves, but the training time and computational resources increase
Solution Approach 1:
The patent applies partial action by using a strategically selected subset of simulation data for training rather than exhaustively using all possible data. The simulation data is designed to cover critical processing scenarios and edge cases that are most relevant to actual manufacturing, allowing the model to achieve high accuracy on essential tasks without the prohibitive computational cost of training on every possible variation
Data Source
AI summary
To provide a model generation method, a recording medium, and an information processing apparatus. The method includes, via a computer: acquiring a plurality of variable groups for constructing a process recipe, state data indicating a state of a processing object before executing a specific step of the process recipe, and target data indicating a target state of the processing object; evaluating the state of the processing object obtained by selecting one variable group from the plurality of variable groups and executing one step characterized by the selected one variable group; and generating a model for constructing the process recipe by reinforcement learning using the acquired state data and target data and a reward determined according to the evaluated state of the processing object.