Graph Neural Network for MEMS Parameter Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In Micro-electro-mechanical systems (MEMS) testing, standard machine learning methods face challenges with sparse and heterogeneous data sets, leading to unreliable and inaccurate predictions, especially when dealing with costly and time-consuming measurements, and are unable to effectively handle missing data or integrate complex relationships between parameters.
Innovation Solution
A method is proposed to train a Graph Neural Network (GNN) with a Heterogeneous Graph Transformer (HGT) architecture to predict second measurement results based on first measurements, constructing graphs that represent relationships between measurements, including spatial and temporal interrelationships, and integrating incomplete data without exclusion or extensive imputation, thereby improving predictive performance on sparse data sets.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If standard machine learning methods are used for predicting measurement results, then the prediction process is simple, but the prediction accuracy and reliability deteriorate when dealing with sparse and heterogeneous data sets
Solution Approach 1:
The patent segments the data processing by dividing the heterogeneous measurement data into structured graph components (nodes representing measurements, edges representing relationships). This segmentation allows the complex prediction task to be broken down into manageable graph traversal and message passing operations, improving prediction accuracy while managing complexity through structured organization.
Solution Approach 2:
The patent transforms the traditional tabular data representation into a graph-based representation, adding a structural dimension to the data. This dimensional change from flat tables to multi-relational graphs enables the model to capture complex interrelationships between measurements, significantly improving prediction accuracy on sparse and heterogeneous data sets.
2Reliability
If more measurements are carried out to improve prediction reliability, then the prediction reliability improves, but the test costs and time consumption increase
Solution Approach 1:
The patent creates a virtual copy of the measurement system through the trained Graph Neural Network model. Once trained on available measurement data, the model can predict unmeasured parameters without physically performing additional measurements, thereby maintaining prediction reliability while eliminating the time and cost of actual additional measurements.
Solution Approach 2:
The patent performs preliminary training of the GNN model using available measurement data before actual prediction tasks. This preliminary action prepares the model to handle sparse and heterogeneous data effectively, enabling reliable predictions without needing to perform additional measurements during the actual testing phase.
3Measurement precision
If incomplete data sets are excluded from analysis to maintain data quality, then the data quality improves, but the quantity of usable data and predictive performance deteriorate
Solution Approach 1:
The patent converts the previously harmful effect of missing data into a benefit by designing a graph-based model that naturally handles incomplete data. The GNN architecture processes only available measurements and propagates information through the graph structure, transforming data sparsity from a liability into a feature that the model can effectively utilize for predictions.
Solution Approach 2:
The patent changes the approach to handling data completeness by shifting from traditional complete-data requirements to a parameterized graph representation where missing values are explicitly modeled as absent edges or nodes. This parameter change allows the system to work with varying degrees of data completeness without compromising prediction quality.
Data Source
AI summary
A computer-implemented method of training a Graph Neural Network for predicting second measurement results of produced products based on received first measurement results is disclosed. The method includes (i) receiving first measurement and second measurement results for a plurality of produced products, (ii) constructing graphs of the first measurements and generating a training data set by assigning the corresponding second measurement of the first measurement to the corresponding graphs, respectively, and (iii) training the Graph Neural Network on the training data set to predict the second measurements based on the graphs.