Neural Network Transistor Compact Modeling for SPICE Simulation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing compact models for semiconductor devices, particularly Field-Effect Transistors (FETs), struggle to accurately capture the electrical characteristics of emerging devices and face challenges in automating model parameter extraction with high accuracy.
Innovation Solution
The development of neural network (NN) models that can be generated from capacitance-voltage (C-V) data, allowing for improved inputs to current-voltage (I-V) based NN models, and enabling the conversion of NNs into circuit simulation code for efficient simulation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If standard FET compact models are used for emerging devices, then model development expertise is required, but the turn-around time becomes long
Solution Approach 1:
The patent replaces the traditional physics-based equation system with a neural network-based data-driven model. This substitution eliminates the need for complex physics equations and manual parameter extraction, enabling automated modeling of emerging devices with long-channel effects while significantly reducing development time and expertise requirements.
Solution Approach 2:
The patent changes the fundamental parameters of the modeling approach by using neural network weights and biases instead of physics-based model parameters. This allows the model to automatically learn device characteristics from simulation or measurement data without requiring manual extraction of physical parameters, thereby reducing both time and expertise requirements.
2Measurement precision
If equation-based models are used, then physics-based accuracy is achieved, but automated parameter extraction with high fitting accuracy remains challenging
Solution Approach 1:
The patent replaces the manual parameter extraction process inherent in equation-based models with an automated neural network training process. The neural network automatically learns optimal parameters from data through backpropagation and optimization algorithms, achieving both high fitting accuracy and full automation without requiring manual intervention.
Solution Approach 2:
The neural network model performs self-service by automatically extracting parameters from simulation or measurement data through the training process. The model self-optimizes its internal parameters (weights and biases) using gradient descent and other optimization techniques, eliminating the need for external manual parameter extraction while achieving high fitting accuracy.
3Measurement precision
If neural network size is increased to improve modeling accuracy, then fitting precision improves, but simulation turn-around time increases
Solution Approach 1:
The patent applies partial action by using a neural network size that is sufficient but not excessive for achieving the required modeling accuracy. The model uses just enough complexity to capture the essential device physics and long-channel effects, avoiding unnecessary computational overhead while maintaining adequate precision for practical applications.
Solution Approach 2:
The patent implements dynamic model complexity adjustment by using separate neural networks for different device types (short-channel and long-channel devices). This allows the system to dynamically select the appropriate model complexity based on the specific device characteristics, optimizing the balance between accuracy and simulation speed for each case.
Data Source
AI summary
A computer implemented method for determining performance of a semiconductor device is provided. The method includes providing training data comprising input state values and training capacitance values to a neural network executing on a computer system; processing the input state values through the neural network to generate modeled charge values; converting the modeled charge values to modeled capacitance values; determining, by the computer system, whether the training capacitance values of the training data are within a threshold value of the modeled capacitance values utilizing a loss function that omits the modeled charge values; and in response to determining that the training capacitance values of the training data are within the threshold value of the modeled capacitance values, converting, by the computer system, the neural network to a circuit simulation code to generate a converted neural network.