Modular Autoencoder Design for Interpretable Process Parameter Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing autoencoder models for manufacturing process parameter estimation are rigid and difficult to interpret, leading to challenges in understanding the flow of information and are prone to overfitting, especially in semiconductor manufacturing where process data complexity is unknown.
Innovation Solution
A modular autoencoder model with adjustable components that allows for easier interpretation and scalability, incorporating known properties during training to reduce bias and avoid overfitting, and enables estimation of parameters from a combination of measurement channels using iterative sub-selection and embedding of domain knowledge.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If a traditional monolithic autoencoder model is used for manufacturing process parameter estimation, then the model can compress high dimensional data to low dimensional representation, but the model becomes rigid and difficult to interpret
Solution Approach 1:
The patent divides the monolithic autoencoder into separate modular components: an encoder module that compresses high-dimensional data to latent space, and a decoder module that reconstructs the data. This segmentation allows independent optimization and interpretation of each module's function, making the overall model more interpretable while maintaining compression capabilities.
Solution Approach 2:
The patent introduces latent space as an intermediary representation between the high-dimensional input data and the output parameters. This latent space acts as a mediator that captures essential information in a compressed form, enabling both data compression and providing a more interpretable representation of the manufacturing process state.
2Reliability
If the autoencoder model complexity is increased to model unknown manufacturing processes, then the model can capture more process characteristics, but the model becomes prone to overfitting noise and unwanted characteristics
Solution Approach 1:
The patent applies partial action by using a fixed, limited number of latent dimensions (e.g., 10-100 dimensions) rather than allowing the model to use all possible dimensions. This partial dimensionality reduction captures the essential process characteristics while filtering out noise and unwanted characteristics, preventing overfitting while maintaining sufficient accuracy.
Solution Approach 2:
The patent changes the parameter of latent space dimensionality from the full high-dimensional space to a reduced low-dimensional space. This parameter change enables the model to generalize better by focusing on the most significant process variations while ignoring noise, thereby improving reliability without excessive complexity.
3Productivity
If the model is trained to compress data to low dimensional representation, then the inference problem is simplified, but the model becomes difficult to extend and adapt to different processes
Solution Approach 1:
The patent segments the model into reusable components (encoder and decoder modules) that can be independently trained and applied to different manufacturing processes. This modular segmentation maintains the computational efficiency of the compressed latent representation while enabling easy extension and adaptation to new processes by simply training new modules rather than redesigning the entire architecture.
Data Source
AI summary
A method for ordering and/or selection of latent elements for modeling low dimensional data within a latent space representation, the low dimensional data being a reduced dimensionality representation of input data as determined by a first model component of a model, comprising the steps of training said model and selecting one of said latent element selections based on said training, said training comprising: reducing a dimensionality of the input data to generate said low dimensional data in said latent space representation; training a second model component of said model for each of one or more latent element selections; and optimizing an approximation of the input data as output by said second model component for each said latent element selection, thereby ranking at least one of said plurality of latent elements in the latent space representation based on a contribution of each latent element to the input data.