Causal Convolution Network for Sparse Semiconductor Process Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current semiconductor manufacturing processes face challenges in accurately representing densely measured parameter values using sparse ADI data, leading to sub-optimal corrections and inefficient throughput due to crosstalk and over-dimensioning of models, as well as limitations in alignment control with sparse measurements.
Innovation Solution
The implementation of a causal convolution neural network that uses input vectors of measured values from previous times to predict parameter values at a later time, allowing for improved configuration and control of semiconductor manufacturing processes by emphasizing temporal dependencies and capturing repeating patterns.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If sparse ADI measurements are used for run-to-run control, then throughput is improved, but manufacturing precision deteriorates due to inaccurate representation of densely measured parameter values
Solution Approach 1:
The system performs preliminary dense measurements at selected time points to train the neural network model, storing these measurements for later use. This preliminary action enables the model to learn accurate representations of process parameters without requiring frequent dense measurements during production, thus maintaining precision while improving throughput.
Solution Approach 2:
The neural network model creates a virtual copy of the dense measurement data by predicting parameter values at times when only sparse measurements are available. This copying mechanism allows the system to use sparse measurements for run-to-run control while maintaining the accuracy equivalent to dense measurements, resolving the contradiction between throughput and precision.
2Manufacturing precision
If dense measurements are performed frequently, then manufacturing precision is improved, but productivity deteriorates due to reduced throughput
Solution Approach 1:
Dense measurements are performed only at selected time points for model training purposes, rather than continuously. This preliminary action captures sufficient information to train the neural network, after which the model can predict parameters using only sparse measurements, thereby maintaining precision while maximizing throughput.
Solution Approach 2:
The system performs dense measurements partially (only at selected time points) rather than excessively (at every time point). This partial action is sufficient to train the model, and the model then compensates for the reduced measurement frequency, achieving the desired precision without the throughput penalty of continuous dense measurements.
3Manufacturing precision
If model parameters are over-dimensioned for sparse data sets, then manufacturing precision may be improved, but device complexity increases
Solution Approach 1:
The neural network model is designed with dynamic capabilities to adapt to varying data densities. The model structure allows it to effectively learn from sparse measurements while maintaining accuracy, avoiding the need for overly complex fixed-dimensional models. This dynamic adaptation resolves the contradiction between precision and complexity.
Data Source
AI summary
A method for configuring a semiconductor manufacturing process, the method comprising: obtaining a plurality of first values of a first parameter based on successive measurements associated with a first operation of a process step in the semiconductor manufacturing process; using a causal convolutional neural network to determine a predicted value of a second parameter based on the first values; and using the predicted value of the second parameter in configuring a subsequent operation of the process step in the semiconductor manufacturing process.