Semiconductor Device I–V Prediction Using Compact-Model GANs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The existing design and simulation processes for semiconductor products are inefficient and costly due to the need for redesign and re-simulation when process levels change, often resulting in inaccurate predictions of device characteristics, which can degrade product performance.
Innovation Solution
A method using deep learning, specifically a generative adversarial network, to predict semiconductor device characteristics by generating simulation current-voltage curves, extracting reference points, and training a model to output prediction curves, with uncertainty data, allowing for efficient and precise prediction of new semiconductor product characteristics.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If design simulation is performed based on changed process level, then prediction accuracy of device characteristics is improved, but time consumption and cost increase enormously
Solution Approach 1:
The patent pre-generates simulation data using compact models for multiple process levels before the actual design process. This preliminary action creates a training dataset that enables the deep learning model to predict device characteristics for new process levels without requiring time-consuming re-simulation, thus resolving the contradiction between accuracy and time consumption
Solution Approach 2:
The patent uses compact models to generate simulated current-voltage curves that copy the behavior of actual semiconductor devices across different process levels. These copied simulation data serve as training samples for the deep learning model, enabling accurate predictions without physical experimentation or full-scale simulation for each new process level
2Productivity
If deep learning model is trained with basic training data, then prediction speed is improved, but prediction accuracy may be insufficient without uncertainty assessment
Solution Approach 1:
The patent implements a feedback mechanism where the deep learning model first generates prediction current-voltage curves, then calculates uncertainty data based on the deviation between predicted and actual simulation data. This uncertainty information feeds back into the system to assess prediction reliability, allowing users to determine whether the rapid prediction results are sufficiently accurate or if further refinement is needed
Solution Approach 2:
The patent performs a two-stage prediction process: first generating prediction curves quickly using the trained deep learning model, then optionally calculating uncertainty data when higher reliability is needed. This partial action approach provides fast predictions for routine cases while enabling more comprehensive analysis when accuracy requirements demand it
3Loss of time
If compact models are used for simulation, then time efficiency is improved, but model accuracy may be insufficient for new process levels
Solution Approach 1:
The patent introduces a deep learning model as an intermediary between compact models and new process level predictions. The deep learning model is trained on simulation data generated by compact models across multiple process levels, learning the relationships and variations. When a new process level is encountered, the deep learning model serves as an intermediary that translates compact model patterns into accurate predictions for the new process level, combining the speed of compact models with the adaptability needed for new processes
Data Source
AI summary
To predict characteristics of a semiconductor device, a simulation current-voltage curve of the semiconductor device is generated using compact models where each compact model generates simulation result data by performing a simulation based on device data. The simulation result data indicate characteristics of semiconductor devices corresponding to the device data. The compact models respectively corresponding to process data and semiconductor products. Simulation reference points on the simulation current-voltage curve are extracted. Basic training data corresponding to a combination of the simulation reference points and the simulation current-voltage curve are generated. A deep learning model is trained based on the basic training data such that the deep learning model outputs a prediction current-voltage curve. A target prediction current-voltage curve is generated based on the deep learning model and target reference points corresponding to the target semiconductor product. The deep learning model is a generative adversarial network.