Substrate Process Defect Prediction Using In-Process Sensor Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The existing substrate processing apparatuses face challenges in determining whether a substrate is defective without undergoing a full defect inspection, leading to potential defective products due to insufficient cleaning.
Innovation Solution
Incorporating sensors to detect physical quantities during polishing, cleaning, and drying, and utilizing a machine learning model to predict defect numbers, sizes, and positions by converting sensor values into feature amounts, allowing for estimation of substrate quality without a defect inspection apparatus.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of energy
If sampling inspection is performed instead of total inspection, then inspection cost is reduced, but the risk of defective products increases
Solution Approach 1:
The patent creates a virtual copy of the defect inspection function by training a machine learning model to predict defect numbers, sizes, and positions based on sensor data from polishing, cleaning, and drying processes. This virtual inspection system replaces or supplements physical defect inspection apparatus, enabling estimation of substrate quality without actual inspection while maintaining high reliability through accurate prediction algorithms
Solution Approach 2:
The patent performs preliminary defect prediction by analyzing process data (sensor values during polishing, cleaning, and drying) before the substrate undergoes actual defect inspection. The machine learning model predicts potential defects based on process conditions, allowing early identification of defective substrates and preventing them from proceeding to costly inspection stages
2Reliability
If total inspection is performed instead of sampling inspection, then defect detection reliability is improved, but inspection cost increases
Solution Approach 1:
The patent creates a virtual copy of the defect inspection function by training a machine learning model to predict defect numbers, sizes, and positions based on sensor data from polishing, cleaning, and drying processes. This virtual inspection system replaces or supplements physical defect inspection apparatus, enabling estimation of substrate quality without actual inspection while maintaining high reliability through accurate prediction algorithms
Solution Approach 2:
The inspection system performs self-service by using process data already collected during normal manufacturing operations (sensor values from polishing, cleaning, and drying) to predict defects. The machine learning model leverages existing process information without requiring additional inspection resources, making the system self-sufficient and cost-effective
3Measurement precision
If defect inspection apparatus is used, then substrate quality is accurately measured, but device complexity and inspection time increase
Solution Approach 1:
The patent replaces the mechanical defect inspection apparatus with a computational system based on machine learning. Instead of using physical inspection equipment to detect defects, the system uses algorithms that process sensor data from manufacturing processes to predict defect characteristics, substituting mechanical inspection with intelligent computational analysis
Solution Approach 2:
The patent introduces machine learning models as intermediaries between process data and defect prediction. The models act as mediators that transform raw sensor values into meaningful quality assessments, eliminating the need for direct physical inspection while maintaining measurement precision through sophisticated data analysis
4Measurement precision
If more sensors and data collection are implemented, then prediction accuracy is improved, but device complexity increases
Solution Approach 1:
The patent makes existing process equipment multi-functional by enabling sensors already installed for process monitoring to also serve defect prediction purposes. The same sensors that monitor polishing, cleaning, and drying processes are utilized to collect data for machine learning models, eliminating the need for dedicated inspection sensors and reducing overall system complexity
Data Source
AI summary
Included are: at least one sensor that detects a physical quantity of an object during polishing and/or during cleaning and/or during drying of a substrate; a conversion section that converts a sensor value during polishing and/or during cleaning and/or during drying detected by the sensor into a feature amount for each processing step with respect to a trained machine learning model; and an inference section that outputs at least one predicted value of a number of defects, a size of a defect, and a position of a defect in a target substrate by inputting target data including the feature amount to the trained machine learning model.