AI Algorithm Data Splitting for Semiconductor Structure Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for measuring semiconductor manufacturing processes face challenges in achieving accuracy, robustness, and precision due to the complexity and cost of securing measurement data, particularly with high-aspect-ratio-contact (HARC) etching, leading to issues with overfitting of AI algorithms when data samples are limited or noisy.
Innovation Solution
A method is provided to load datasets representing semiconductor spectra and structures, calculate an out-of-distribution (OOD) index, perform data split by cluster sampling, and select an optimal AI algorithm trained on learning datasets to prevent overfitting and ensure robustness even with a small amount of data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If measurement tools are used for semiconductor manufacturing processes, then measurement data can be obtained, but measurement accuracy and precision deteriorate due to process complexity and tool limitations
Solution Approach 1:
The patent introduces an intermediary AI model that acts as a mediator between spectrum data and structure information. Instead of directly measuring complex semiconductor structures with limited measurement tools, the system uses spectrum data as an intermediary representation that can be processed by the AI model to infer structural information, thereby overcoming the precision limitations of direct measurement tools
Solution Approach 2:
The patent replaces the mechanical measurement system with an AI-based predictive system. Instead of relying on physical measurement tools that struggle with complex HARC etching structures, the system substitutes a computational approach using AI models trained on spectrum data to predict structural characteristics, eliminating the need for direct physical measurement of complex structures
2Loss of time
If AI algorithms are trained with limited data samples, then training time is reduced, but overfitting occurs leading to poor generalization performance
Solution Approach 1:
The patent applies preliminary action by performing data augmentation and synthetic data generation before training the AI model. Multiple synthetic datasets are generated with varying conditions and noise levels in advance, allowing the model to be trained on diverse data distributions without requiring extensive collection of real measurement data, thus preventing overfitting while maintaining efficient training
Solution Approach 2:
The patent changes data parameters by generating synthetic spectrum data with varied noise levels, scaling factors, and structural parameters. This parameter variation in synthetic data allows the AI model to learn robust features that generalize well across different conditions, preventing overfitting to specific data characteristics while reducing dependency on large volumes of real data
3Reliability
If measurement processes are made more robust to handle complex structures, then measurement reliability improves, but process complexity and cost increase
Solution Approach 1:
The patent extracts the complex measurement process and replaces it with a simplified spectral analysis approach. Instead of using complex measurement tools and procedures to directly characterize HARC etching structures, the system extracts spectrum data that contains structural information and processes it through AI models, thereby achieving robust structural prediction without the complexity of advanced measurement systems
Data Source
AI summary
Provided are a method of providing an artificial intelligence (AI) algorithm, an operation method of an AI algorithm, an electronic device, a recording medium, and a computer program. The method of providing the AI algorithm includes loading data sets with respect to a spectrum of a semiconductor and a structure of the semiconductor, calculating an out of distribution (OOD) index with respect to the spectrum of the semiconductor, performing data split by clustering sampling the data sets into at least one learning data set with respect to OOD indexes according to semiconductors, and providing an optimal AI algorithm from among a plurality of AI algorithms that have learned the at least one learning data set.