Machine Learning Silicon Density Prediction for Semiconductor Metal Layers
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for calculating silicon density in metal layers of semi-conductor chips are time-consuming and inefficient, consuming 50%-60% of parasitic extraction runtime due to the need for detailed etching calculations.
Innovation Solution
The use of machine learning, specifically a model containing a convolution neural network (CNN) followed by a deep neural network (DNN), to predict silicon density for a metal layer based on image features and drawn density, reducing the need for detailed etching calculations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If detailed etching calculations are performed to accurately determine silicon density, then manufacturing precision is improved, but productivity deteriorates due to increased runtime
Solution Approach 1:
The patent creates a machine learning model that learns from detailed etching calculation results (training data) and then uses this learned model to predict silicon density without performing full etching calculations. The model captures the relationship between drawn density and actual silicon density through training on representative samples, enabling accurate predictions at a fraction of the computational cost.
Solution Approach 2:
The patent transforms the complex multi-parameter etching calculation problem into a simpler prediction problem by training a machine learning model on input-output pairs. The model learns to map drawn density parameters to actual silicon density outcomes, effectively changing the computational approach from physics-based simulation to data-driven prediction.
2Measurement precision
If detailed etching calculations are performed for all geometries, then measurement precision is improved, but loss of time increases significantly
Solution Approach 1:
The patent performs preliminary etching calculations on a training set of geometries to establish the relationship between drawn and actual density. This preliminary action creates a trained machine learning model that can then rapidly predict silicon density for production cases without repeating the full etching calculation process.
Solution Approach 2:
The patent uses training data from detailed etching calculations to create a predictive model that copies the essential relationships without requiring the full computational process for each new geometry. The model captures the etching behavior patterns and applies them efficiently to new designs.
Data Source
AI summary
A specification for a semi-conductor chip is received. The specification specifies a set of photomasks associated with a metal layer of the semi-conductor chip. Multiple portions of an area of the metal layer are identified. A respective image is generated for each portion of the area based on the photomasks. A respective drawn density of metal wires for each portion of the area is calculated. A trained machine learning model is invoked to predict a respective silicon density of metal wires for each respective portion of the area based on an image and a drawn density for the respective portion of the area. A silicon density for the area of the metal layer is calculated based on a combination of predicted silicon densities for the multiple portions of the area. The combination is based on an average value of the predicted silicon densities for the multiple portions of the area.