Transfer Learning for Semiconductor ML Model Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current semiconductor manufacturing processes face challenges in efficiently training machine learning models due to limited training samples, requiring ad hoc and lossy methods that do not fully utilize metrology and inspection equipment capabilities, and are slow with rigorous process simulation designs that often lack necessary optical proximity correction data.
Innovation Solution
A system and method for training a machine learning model using information from nominal instances of specimens, with re-training on non-nominal instances through transfer learning, employing optical and electron beam imaging subsystems to acquire and process data for efficient simulations in inspection, metrology, and defect review.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If rigorous process simulation design of experiments (DOES) through process window is employed, then measurement precision is improved, but training time and productivity deteriorate significantly
Solution Approach 1:
The patent pre-generates a comprehensive library of simulated wafer images covering the full process window (focus, exposure, dose variations) before inspection. This preliminary simulation creates training data in advance, eliminating the need for time-consuming rigorous DOES during actual model training while maintaining measurement precision through the pre-computed process window coverage.
Solution Approach 2:
The patent uses simulated wafer images as copies of real wafer appearances to create training datasets. These synthetic images replicate the visual characteristics of actual wafers under various process conditions, allowing the machine learning model to learn from numerous simulated examples without requiring equivalent numbers of real physical wafers for each process condition.
2Reliability
If transfer learning is used to re-train the machine learning model with non-nominal instances, then classification accuracy is improved, but training time increases
Solution Approach 1:
The patent pre-generates both nominal and non-nominal simulated wafer images in advance, organizing them into training datasets before the transfer learning process. This preliminary preparation of diverse training samples (including defective patterns, process variations, and edge cases) enables the model to achieve high classification accuracy during transfer learning without requiring extensive real-time data collection or processing.
3Productivity
If ad hoc and lossy methods are used for model training, then training speed is improved, but the ability to utilize metrology and inspection equipment capabilities deteriorates
Solution Approach 1:
The patent creates comprehensive simulated wafer image libraries that copy and replicate the full range of appearances that real wafers can exhibit under various process conditions. This simulated data copying preserves rich information about process variations, defect patterns, and measurement characteristics that would be lost in ad hoc training methods, while still enabling fast training through synthetic data generation rather than slow real data collection.
Data Source
AI summary
Methods and systems for accelerated training of a machine learning based model for semiconductor applications are provided. One method for training a machine learning based model includes acquiring information for non-nominal instances of specimen(s) on which a process is performed. The machine learning based model is configured for performing simulation(s) for the specimens. The machine learning based model is trained with only information for nominal instances of additional specimen(s). The method also includes re-training the machine learning based model with the information for the non-nominal instances of the specimen(s) thereby performing transfer learning of the information for the non-nominal instances of the specimen(s) to the machine learning based model.