Learning Device for Semiconductor Simulation Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Semiconductor manufacturing processes are complex and cannot be accurately represented by physical models, limiting the accuracy of simulations, which motivates the exploration of machine learning-based models as an alternative to improve simulation accuracy.
Innovation Solution
A method involving a learning device that processes non-processed object image data and process-related data to infer processed object image data using a learned model, trained to reflect the effects of manufacturing processes, thereby enhancing simulation accuracy by correlating with events in the semiconductor manufacturing process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If physical models are used to simulate semiconductor manufacturing processes, then the simulation framework is established, but the simulation accuracy is limited due to complex process behaviors that cannot be represented by physical models
Solution Approach 1:
The patent replaces physical model-based simulation with machine learning model-based simulation. The learning device trains a machine learning model using training data consisting of process parameters and inspection results, then uses this trained model to simulate manufacturing processes. This substitution eliminates the need to define complex physical equations while achieving higher simulation accuracy that captures real process behaviors.
2Manufacturing precision
If machine learning models are used instead of physical models, then simulation accuracy is improved, but the complexity of model training and data processing increases
Solution Approach 1:
The patent performs preliminary action by collecting and preparing training data before model deployment. The learning device collects process parameters and inspection results, preprocesses this data, and trains the machine learning model in advance. This preliminary training phase enables the model to learn complex process behaviors, after which the trained model can be used for simulation without requiring complex real-time processing.
Solution Approach 2:
The machine learning model serves itself by automatically learning process behaviors from training data without requiring manual definition of physical equations. The model self-adjusts its parameters during training to minimize prediction errors, and once trained, it autonomously performs simulation tasks based on input process parameters, eliminating the need for continuous manual intervention.
3Productivity
If physical models are used for simulation, then the simulation framework is established, but the simulation time is extended due to the complexity of physical equations
Solution Approach 1:
The patent substitutes computational physics calculations with pre-trained machine learning model inference. Instead of solving complex physical equations in real-time, the system uses the trained model to directly predict inspection results from process parameters. This substitution dramatically reduces simulation time while maintaining or improving accuracy, as the model has already learned the underlying patterns during the training phase.
Data Source
AI summary
With respect to a method performed by at least one processor, the method includes obtaining, by the at least one processor, data related to a first process for a first object, obtaining, by the at least one processor, non-processed object data of the first object, generating, by the at least one processor, first data including the data related to the first process for the first object and the non-processed object data of the first object, and adjusting a second process for a second object based on the first data.