Learning Model for Semiconductor Simulation Speed
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Semiconductor manufacturing processes require extensive time for simulations using physical models, prompting the need for alternative methods that can efficiently replicate simulation results, such as those achieved by machine learning models.
Innovation Solution
A learning model is trained to process non-processed image data and parameter data, using a U-shaped convolutional neural network (UNET) to generate simulated results that match those produced by physical model simulators, thereby reducing simulation time and maintaining accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If physical models with repeated trials are used for semiconductor manufacturing simulations, then simulation accuracy is improved, but simulation time increases significantly
Solution Approach 1:
The patent creates a learned model that copies the simulation results of physical models through machine learning. The learned model is trained using input-output pairs from physical model simulations, enabling it to reproduce simulation results without requiring repeated physical model trials, thus achieving both accuracy and speed.
Solution Approach 2:
The patent performs preliminary training of the learned model using physical model simulation data before actual manufacturing simulations. This preliminary action creates a pre-trained model that can quickly predict results without requiring repeated physical model simulations during production, reducing simulation time while maintaining accuracy.
2Reliability
If traditional physical model simulators are used, then simulation accuracy is maintained, but productivity decreases due to extensive simulation time
Solution Approach 1:
The learned model copies the functional behavior of physical model simulators by training on their output data. This allows the system to maintain simulation accuracy while achieving faster execution speeds, thereby improving productivity without sacrificing reliability.
Solution Approach 2:
The patent changes the computational parameters from physical model-based calculations to machine learning model predictions. This parameter change transforms the simulation process from computationally intensive physical calculations to faster neural network inference, improving productivity while maintaining accuracy through proper training.
Data Source
AI summary
With respect to an inference method performed by at least one processor, the method includes inputting, by the at least one processor, into a learned model, second non-processed image data and second parameter data of a simulator, and inferring, by the at least one processor using the learned model, second processed image data. The learned model has been trained so that first processed image data, obtained as an output in response to first non-processed image data and first parameter data of the simulator for the first non-processed image data being input, approaches first simulator processed image data, obtained as a result of the simulator for the first non-processed image data by using the first parameter data.