Semiconductor Device Modeling with Reinforcement Learning Parameter Fitting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The existing design technology co-optimization (DTCO) methods for semiconductor products are inefficient, requiring significant time and cost to update device models when target process levels change, and often result in inaccurate predictions of semiconductor device characteristics.
Innovation Solution
A method using reinforcement learning to optimize device model parameters, which involves receiving measurement data, selecting target parameters, determining local minimum values, and improving parameter values to enhance prediction accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional device model updating methods are used when target process level changes, then the device model can be updated, but it requires enormous amount of time and impractical effort to perform new simulations
Solution Approach 1:
The patent performs preliminary actions by pre-calculating and storing device characteristics data across multiple process levels before actual design simulation. When a process level change occurs, the system can directly retrieve pre-computed data instead of performing time-consuming simulations, thus resolving the contradiction between model accuracy and simulation time.
Solution Approach 2:
The patent creates copies of device model data across different process levels. By generating and storing device characteristics data for multiple process levels in advance, the system can switch between process levels by retrieving pre-computed copies rather than performing new simulations, eliminating the time loss while maintaining accuracy.
2Reliability
If traditional device model updating methods are used when target process level changes, then the device model can be updated, but extra cost is incurred to the manufacture
Solution Approach 1:
The system performs device model updates and simulations in advance for multiple process levels, storing the results for later retrieval. This preliminary action eliminates the need for costly re-simulations when process levels change, reducing manufacturing costs while maintaining model accuracy.
Solution Approach 2:
The patent creates and stores copies of device model data for different process levels. When manufacturing requires a different process level, the system retrieves pre-computed copies instead of performing new simulations, thereby reducing manufacturing costs while preserving model reliability.
3Productivity
If low accuracy model is used for simulation, then simulation time is reduced, but the performance of designed and manufactured product is degraded due to inaccurate prediction
Solution Approach 1:
The system performs high-accuracy simulations in advance and stores the results. During actual design work, it retrieves pre-computed high-accuracy data instantly, achieving both fast simulation speed and high prediction accuracy simultaneously, thus resolving the contradiction between productivity and measurement precision.
Solution Approach 2:
The patent creates copies of high-accuracy device characteristics data for multiple process levels and stores them for rapid retrieval. This allows the system to provide accurate predictions quickly without performing time-consuming simulations, resolving the contradiction between simulation speed and prediction accuracy.
Data Source
AI summary
Measurement data are produced by measuring characteristics of a semiconductor device. Target parameters are selected among a plurality of parameters of a device model where the device model is configured to perform a simulation based on device data and output simulation result data indicating the characteristics of the semiconductor device. Initial value sets corresponding to different combinations of initial values of the target parameters are selected. Local minimum values are determined based on reinforcement learning. Each local minimum value corresponds to a minimum value of a difference between the measurement data and the simulation result data with respect to each initial value set. Optimal values of the target parameters are determined based on the plurality of local minimum values. The device model capable of precisely predicting characteristics of the semiconductor device is generated by determining the parameters of the device model using the optimization scheme based on the reinforcement learning.