Deep Neural Network Semiconductor Modeling with Transformed Targets
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current semiconductor device models, particularly those using artificial neural networks, face challenges in accurately modeling currents over a wide range, including the sub-threshold region, due to over-fitting and poor monotonicity, which affects the accuracy of conductance and transconductance values.
Innovation Solution
The use of a deep neural network that targets a transformed version of the semiconductor drain current, processed through a pre-processor and adjusted using a loss function that includes a complexity cost to prevent over-fitting, allows for more precise modeling of sub-threshold conditions and improved accuracy across a wider range of input voltages.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a neural network is trained to model drain current directly, then the model can capture complex device behaviors, but the model suffers from over-fitting and poor monotonicity leading to inaccurate conductance and transconductance values
Solution Approach 1:
The patent introduces transformed target variables (logarithm of drain current, gate voltage, and drain voltage) as intermediaries in the training process. Instead of directly modeling the raw drain current, the neural network models the transformed variables, which serve as mediators to improve the quality of training data and guide the network toward better convergence and monotonicity in the final current predictions
Solution Approach 2:
The patent applies parameter transformation by taking the logarithm of the drain current, gate voltage, and drain voltage as target variables for training. This parameter change transforms the training objective from directly predicting current to predicting the log of current, which stabilizes the training process and improves the accuracy of derived parameters like conductance and transconductance
2Adaptability or versatility
If the neural network models a wide range of currents including sub-threshold region, then the model coverage is improved, but the accuracy deteriorates due to over-fitting in certain regions
Solution Approach 1:
The patent transforms the target parameters by applying logarithmic functions to the drain current, gate voltage, and drain voltage. This parameter transformation equalizes the scale across different operating regions including sub-threshold, threshold, and strong inversion regions, allowing the neural network to learn patterns across the full range without being dominated by large current values in strong inversion
Solution Approach 2:
The patent uses a loss function that computes the mean squared error between the transformed target variables and the neural network predictions. This feedback mechanism guides the training process to minimize errors across all operating regions, with the transformed variables ensuring that errors in low-current sub-threshold regions are weighted appropriately alongside high-current regions
Data Source
AI summary
A deep neural network models semiconductor devices. Measurements of test transistors are gathered into training data including gate and drain voltages and transistor width and length, and target data such as the drain current measured under the input conditions. The training data is converted by an input pre-processor that can apply logarithms of the inputs or perform a Principal Component Analysis (PCA). Rather than use measured drain current as the target when training the deep neural network, a target transformer transforms the drain current into a transformed drain current, such as a derivative of the drain current with respect to gate or drain voltages, or a logarithm of the derivative. Weights in the deep neural network are adjusted during training by comparing the deep neural network's output to the transformed drain current and generating a loss function that is minimized over the training data.