Modular Autoencoder for Semiconductor Parameter Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current data-driven inference approaches for semiconductor metrology operations face challenges in efficiently selecting the best channels for measurement due to time-consuming brute force methods, which can lead to overfitting and introduce channel-specific biases, limiting their predictive consistency and accuracy in estimating semiconductor manufacturing process parameters.
Innovation Solution
A modular autoencoder model is employed to estimate parameters from a combination of available channels by iteratively varying and training on subsets of channels, ensuring the model remains predictive across any channel combination and avoids channel-specific biases, while also embedding known properties during training to reduce inference biases.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If brute force methods are used to select the best channels for measurement, then comprehensive channel evaluation is achieved, but measurement time and computational resources increase significantly
Solution Approach 1:
The patent segments the channel selection process into multiple iterations, where in each iteration only a subset of channels is trained and evaluated. This divides the comprehensive channel evaluation into manageable segments, reducing the computational burden and measurement time for each individual evaluation while still achieving comprehensive selection across all iterations.
Solution Approach 2:
The patent implements partial action by training on subsets of channels rather than all channels simultaneously. This partial training approach allows for faster individual iterations while the cumulative effect across multiple iterations with different subsets achieves comprehensive channel evaluation, effectively reducing overall measurement time.
2Reliability
If training is performed on all channel combinations, then complete channel evaluation is achieved, but the model becomes biased toward specific channels
Solution Approach 1:
The patent merges the results from multiple iterations, each trained on different channel subsets. By combining the knowledge gained from various subsets across iterations, the final model achieves comprehensive channel evaluation while distributing the training load, thereby preventing over-reliance on any single channel combination and reducing channel-specific biases.
Solution Approach 2:
The patent introduces dynamics by varying the channel subsets across different iterations rather than using a fixed channel combination. This dynamic approach ensures that the model learns from diverse channel configurations, improving predictive consistency across different channel combinations while preventing the model from becoming biased toward any specific static channel set.
3Measurement precision
If complex monolithic autoencoder models are used, then modeling accuracy is improved, but model interpretability and ease of extension decrease
Solution Approach 1:
The patent segments the autoencoder model into a modular architecture with separate encoder and decoder components. This segmentation improves interpretability by allowing independent analysis of the encoding and decoding processes, while still maintaining the capability to achieve high parameter estimation accuracy through the coordinated work of these modular components.
Solution Approach 2:
The patent performs preliminary dimensionality reduction through the encoder component before the main parameter estimation task. This preliminary action simplifies the subsequent processing by working with compressed latent representations, making the overall system more interpretable while maintaining accuracy by preserving essential information in the compressed form.
Data Source
AI summary
A modular autoencoder model is described. The modular autoencoder model comprises input models configured to process one or more inputs to a first level of dimensionality suitable for combination with other inputs: a common model configured to: reduce a dimensionality of combined processed inputs to generate low dimensional data in a latent space; and expand the low dimensional data in the latent space into one or more expanded versions of the one or more inputs suitable for generating one or more different outputs; output models configured to use the one or more expanded versions of the one or more inputs to generate the one or more different outputs, the one or more different outputs being approximations of the one or more inputs; and a prediction model configured to estimate one or more parameters based on the low dimensional data in the latent space.