Virtual Metrology Models for Sparse Semiconductor Process Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Modern manufacturing processes, particularly in semiconductor manufacturing, face challenges due to the limited amount of sparse measurement data, which hinders effective process control and monitoring. Existing systems struggle to accurately model data drifts, shifts, and hierarchical data structures, leading to sub-optimal control and increased costs.
Innovation Solution
The system employs a data and machine learning layer that includes a data collector, dataset generator, model management module, and inference module. This configuration allows for the creation of customized datasets and the building, training, and updating of machine learning models. The aggregated adaptive online model (AggAOM) is used to leverage commonalities across equipment and chambers, addressing data scarcity and sparsity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If sparse measurement data is used for process control and monitoring, then data collection burden is reduced, but prediction accuracy and manufacturing precision deteriorate
Solution Approach 1:
The patent creates virtual copies of physical measurements through machine learning models. The system trains models on available sparse measurement data to generate virtual measurement data that replicates the characteristics of actual measurements, thereby expanding the effective data sample size without additional physical measurement burden
Solution Approach 2:
The system performs preliminary data preparation and model training in advance. Historical measurement data is collected and used to train virtual metrology models before actual process control is needed, so that when sparse measurements are available, the pre-trained models can immediately generate accurate predictions without real-time computation delays
2Manufacturing precision
If more measurement data is collected to improve prediction accuracy, then manufacturing precision improves, but data processing complexity and cost increase
Solution Approach 1:
Instead of collecting more physical measurements which would require more sensors and processing infrastructure, the system creates virtual copies of measurements through ML models, achieving the same data volume and accuracy goals with minimal additional hardware or processing complexity
Solution Approach 2:
The virtual metrology system serves multiple functions: it generates predictive measurements, performs anomaly detection, and provides process optimization insights, all through the same ML model infrastructure, thereby reducing overall system complexity compared to implementing separate systems for each function
3Productivity
If virtual metrology systems are deployed to improve productivity and quality, then manufacturing efficiency improves, but model accuracy and scalability deteriorate due to data drifts and shifts
Solution Approach 1:
The patent implements dynamic model updating mechanisms that continuously adapt to changing process conditions. The system monitors data drifts and shifts in real-time and automatically retraines or adjusts the ML models to maintain accuracy, transforming static models into dynamic systems that evolve with the manufacturing process
Solution Approach 2:
The system incorporates feedback loops where prediction results are compared against actual measurements when available. This feedback is used to continuously refine and improve the virtual metrology models, ensuring they remain accurate even as process conditions change over time
Data Source
Figure 1A
Figure 1B
Figure 1C
AI summary
Described are systems and methods for optimizing process control or monitoring of manufacturing processes in manufacturing environments. Systems and methods can generate predictions or recommendations for process variables, target properties, or root causes of anomalies. Systems can include data and machine learning layers that can include: a data collector configured to receive data from the client application layer; a dataset generator configured to enable a user to create customized datasets from the data; a model management module configured to enable the user to build, train and/or update machine learning models; and an inference module configured to use the machine learning models for generating predictions or recommendations. Machine learning models can include aggregated adaptive online models (AggAOM) for generating predictions with scarce or sparse data.