Semiconductor Measurement Models Using Historical Data for Sparse References
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning (ML) based metrology methods for advanced semiconductor structures face inaccuracies due to limited and sparse reference measurement data, leading to poor characterization of complex structures with diverse geometries and materials, and synthetic data sets fail to capture pre-process variations effectively.
Innovation Solution
Employ historical measurement data and prior state data to train a present state, ML based measurement model, leveraging correlations between different design and process revisions to improve measurement accuracy and robustness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If machine learning based measurement models are used for advanced semiconductor structures, then measurement speed and throughput are improved, but measurement accuracy deteriorates due to limited and sparse reference measurement data
Solution Approach 1:
The patent applies preliminary action by collecting and storing reference measurement data from multiple prior process states and historical process revisions before training the ML model. This pre-collected data serves as a comprehensive training dataset that enables the model to achieve high measurement accuracy without requiring extensive reference measurements during actual production, thus maintaining both speed and accuracy
Solution Approach 2:
The patent uses copying by creating synthetic training data through simulations of complex semiconductor structures. These simulated structures replicate real device geometries and materials, providing abundant training examples that would be difficult or time-consuming to obtain through actual reference measurements, thereby improving model accuracy while maintaining measurement throughput
2Measurement precision
If more reference measurement data is collected to improve measurement accuracy, then measurement precision is improved, but measurement time and data collection time increase
Solution Approach 1:
The patent performs preliminary data collection by gathering reference measurements from multiple prior process states and historical revisions in advance. This pre-collected data is stored and used to train the ML model, eliminating the need to collect extensive reference data during production measurements, thus achieving high accuracy without time loss
Solution Approach 2:
The patent implements continuous data collection across multiple process states and revisions, accumulating reference data continuously as it becomes available. This continuous accumulation builds a comprehensive training dataset over time, enabling the ML model to achieve high measurement accuracy while the actual measurement process remains rapid and efficient
3Reliability
If complex three-dimensional semiconductor structures are fabricated to meet device specifications, then device performance is improved, but characterization difficulty increases
Solution Approach 1:
The patent creates simulated copies of complex three-dimensional semiconductor structures through computational modeling. These simulated structures replicate the geometries and materials of actual devices, providing training data for the ML model without requiring physical reference measurements. This approach enables accurate characterization of complex structures while avoiding the difficulties and time requirements of physical reference measurement
Solution Approach 2:
The patent transforms the characterization problem by changing parameters from direct physical measurement to ML-based prediction. The ML model learns to predict structure parameters from measurement data by training on simulated and historical data, enabling accurate characterization of complex three-dimensional structures without requiring complex physical measurement procedures
4Measurement precision
If physical model based metrology methods are used with regression and simulation, then measurement accuracy is improved, but computation time and overall time to generate results increases significantly
Solution Approach 1:
The patent replaces complex physical model-based methods with a machine learning model that has been trained on simulated data. The ML model learns the relationship between measurement data and structure parameters from training examples, enabling rapid prediction without requiring time-consuming regression analysis or simulation during actual measurements, thus achieving both accuracy and speed
Solution Approach 2:
The patent substitutes the mechanical computation process of physical model-based metrology (regression, simulation, iteration) with a machine learning inference process. The ML model performs rapid pattern recognition and parameter prediction based on learned relationships, replacing the computationally intensive physical model approach with a faster data-driven approach that maintains measurement accuracy
Data Source
AI summary
Methods and systems for using historical measurement data to train a present state, machine learning (ML) based measurement model are described herein. This approach takes advantage of the correlation between structural characteristics of measured samples fabricated in accordance with different design revisions, process revisions, or both. In one aspect, a present state, ML based measurement model is trained using training data associated with measurements of a plurality of instances of a current version of a semiconductor structure in a present state of a semiconductor process flow and training data associated with measurements of a plurality of instances of a historical version of the semiconductor structure in the present state of the semiconductor process flow. In some examples, training data also includes prior state measurement data. Historical training data, prior state training data, or both, may be derived from actual reference measurements, in-line, production measurements, or both.