Robust Machine Learning Model for Semiconductor Process Drift
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models in semiconductor manufacturing face challenges due to time-dependent drift in sensor measurements, which can lead to inaccurate predictions, as they assume similarity between training and test data sets, failing to account for environmental and manufacturing variations.
Innovation Solution
The approach involves modeling temporal dependencies of sensor measurements using techniques like linear regression, time series regression, and Kalman Filters to generate robust machine learning models that are insensitive to time-dependent shifts, ensuring predictions are reasonable and accurate for future production runs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional machine learning models are used assuming training and test data similarity, then model construction is simple and fast, but prediction accuracy deteriorates due to time-dependent drift in sensor measurements
Solution Approach 1:
The patent segments the machine learning model into multiple components: a base ML model for capturing general patterns and additional time-dependent models (such as autoregressive models or trend adjustment models) for capturing temporal drift. This segmentation allows each component to specialize in specific aspects, improving overall prediction accuracy while keeping individual components manageable in complexity
Solution Approach 2:
The patent introduces dynamic elements to the ML model by incorporating time-dependent parameters and adaptive mechanisms that allow the model to adjust to changing conditions over time. This includes using rolling window approaches, adaptive learning rates, and time-varying parameters that enable the model to track temporal drift without requiring complete model reconstruction
2Reliability
If robust design methods are applied to achieve insensitivity to noise variations, then reliability under environmental variation improves, but nominal performance deteriorates
Solution Approach 1:
The patent employs parameter changes by adjusting model parameters dynamically based on temporal patterns detected in the data. Instead of using fixed robust parameters, the system adapts parameters such as sensitivity weights, time constants, and drift compensation factors to maintain optimal performance across different operating conditions, thereby preserving nominal performance while achieving robustness
3Reliability
If time-dependent drift in sensor measurements is not accounted for, then model construction remains simple, but prediction reliability deteriorates for future production runs
Solution Approach 1:
The patent applies preliminary action by pre-processing the training data to identify and characterize temporal drift patterns before model construction. This includes performing initial trend analysis, seasonal decomposition, and drift detection on historical data to establish baseline temporal characteristics. These preliminary insights are then used to guide the construction of drift-compensated models, improving future prediction reliability without requiring overly complex architectures
Data Source
AI summary
Robust machine learning predictions. Temporal dependencies of process targets for different machine learning models can be captured and evaluated for the impact on process performance for target. The most robust of these different models is selected for deployment based on minimizing variance for the desired performance characteristic.