Semiconductor Compact Modeling Using Mixture of Experts Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional neural network-based compact modeling for semiconductor devices requires extensive training data and time, as it attempts to model various device operation regions simultaneously, leading to inefficiencies in generating compact models.
Innovation Solution
The use of multiple specialized artificial neural networks, each focused on specific semiconductor device operation regions, such as short channel effects, on/off states, and current estimation in different regions, through a mixture of experts (MoE) stages, reducing the need for extensive training data and time.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If one neural network is used to model all device operation regions simultaneously, then comprehensive device characteristics can be captured, but training time and data requirements increase significantly
Solution Approach 1:
The patent divides the single neural network into multiple specialized neural networks, each trained to model specific device operation regions (cutoff region, linear region, saturation region). This segmentation allows each network to focus on particular characteristics, reducing the overall training time and data requirements while maintaining comprehensive coverage of device behavior through selective combination of specialized networks.
2Adaptability or versatility
If one neural network is used to model all device operation regions simultaneously, then comprehensive device characteristics can be captured, but training data requirements increase significantly
Solution Approach 1:
The patent segments the modeling task into multiple specialized neural networks, each requiring less training data for its specific operation region. By dividing the comprehensive modeling task into smaller specialized tasks, the total training data requirement is reduced while still achieving comprehensive device characteristic coverage through the combination of specialized networks.
Solution Approach 2:
Each neural network is specialized to model specific local characteristics of device operation regions rather than attempting to model all regions uniformly. This local quality approach allows each network to achieve high accuracy with less data by focusing on the specific characteristics of its designated operation region.
3Loss of time
If multiple specialized neural networks are used for each operation region, then training time is reduced, but system complexity increases
Solution Approach 1:
The patent segments the complex modeling task into multiple simpler specialized networks, reducing training time for each. The system manages this complexity through a structured framework that selects and combines the appropriate specialized networks based on the input conditions, transforming complexity from training to runtime execution.
Solution Approach 2:
The system dynamically selects which specialized neural networks to use based on the input parameters and operating conditions. This dynamic selection approach allows the system to manage complexity by activating only the necessary specialized networks for each specific case, rather than requiring all networks to be simultaneously active and managed.
4Productivity
If multiple specialized neural networks are used for each operation region, then compact model generation efficiency is improved, but model structure complexity increases
Solution Approach 1:
The patent segments the compact model generation process into multiple specialized neural networks, each handling specific operation regions. This segmentation improves generation efficiency by allowing parallel training and specialized optimization, while the modular structure makes the overall system more manageable despite the increased number of components.
Data Source
AI summary
A method of semiconductor device compact modeling using multiple specialized artificial neural networks for each semiconductor device operation region. The method can include applying channel width data, channel length data, or temperature data of the semiconductor device to a first mixture of experts (MoE) stage to generate a first MoE stage output including first information on characteristics of the semiconductor device according to presence or absence of a short channel effect of the semiconductor device. The method can also include applying the first MoE stage output and gate-source voltage data to a second MoE stage to generate a second MoE stage output including second information on the characteristics of the semiconductor device according to an on state or off state of the semiconductor device.