FDC Model Retraining for Semiconductor Waveform Drift
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In semiconductor manufacturing, the increasing complexity and precision of processes require frequent adjustments to analysis models for accurate monitoring, which is labor-intensive and often not possible in real time without manual intervention.
Innovation Solution
An automated operation method and device for a failure detection and classification (FDC) model that continuously monitors raw traces for changes, segments data into windows, calculates indicators, and retraining the model based on these indicators to adapt to new waveforms and correct offset forecasts.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual monitoring and adjustment of the analysis model is performed, then the model can be adjusted when needed, but considerable manpower is consumed and real-time correction is not possible
Solution Approach 1:
The system automatically monitors raw traces, detects waveform changes, segments data into windows, calculates indicators, and retrains the FDC model without human intervention. The model self-corrects by automatically identifying when retraining is needed and executing the retraining process, eliminating the need for manual monitoring and adjustment while maintaining high accuracy
Solution Approach 2:
The system continuously monitors raw traces and compares them against the current FDC model performance. When waveform changes are detected or model accuracy degrades, the system automatically triggers retraining with new data windows and indicators, creating a closed-loop feedback mechanism that maintains model accuracy in real-time without manual intervention
2Reliability
If the analysis model is frequently adjusted and retrained to handle increased product complexity and process precision, then monitoring accuracy is improved, but the process becomes more complex and labor-intensive
Solution Approach 1:
The system automatically segments raw traces into multiple windows based on detected waveform changes. Each window represents a specific time period or process stage, allowing the model to adapt to different process conditions independently. This segmentation enables targeted retraining on specific process phases rather than requiring complete model redesign
Solution Approach 2:
The system dynamically adjusts the FDC model based on real-time detection of waveform changes in raw traces. When process conditions change or new failure modes emerge, the system automatically updates the model with new data windows and indicators, making the monitoring system adaptive to increasing product complexity and process precision requirements
3Reliability
If manual monitoring is used to detect when the analysis model needs adjustment, then model performance can be maintained, but real-time correction is not achievable
Solution Approach 1:
The system continuously monitors raw traces in real-time, constantly comparing actual process data against the FDC model predictions. This continuous monitoring enables immediate detection of waveform changes or model drift, allowing for real-time model correction rather than periodic manual checks, thus eliminating time delays in maintaining monitoring accuracy
Data Source
AI summary
An operation method and an operation device of a failure detection and classification (FDC) model are provided. The operation method of the FDC model includes the following steps. A plurality of raw traces are continuously obtained. If the raw traces have started to be changed from the first waveform to the second waveform, whether at least N pieces in the race traces have been changed to the second waveform is determined. If at least N pieces in the raw traces have been changed to the second waveform, the raw traces which have been changed to the second waveform are automatically segmented to obtain several windows. An algorithm is automatically set for each of the windows. Through each of the algorithms, an indicator of each of the windows is obtained. The FDC model is retrained based on these indicators.