Semiconductor Characteristic Prediction Using Compact-Model Deep Learning
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 frequent adjustments in response to changes in manufacturing processes, leading to inaccurate predictions of device characteristics and potential performance degradation.
Innovation Solution
A method using deep learning models trained with basic training data from compact models to predict semiconductor device characteristics, incorporating uncertainty analysis for improved accuracy and efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If traditional simulation-based design methods are used, then manufacturing accuracy can be maintained, but design time and cost increase significantly
Solution Approach 1:
The patent pre-trains deep learning models using historical simulation data from compact models before actual device design. This preliminary training creates a predictive system that can quickly estimate device characteristics without requiring full re-simulation, thus reducing design time while maintaining accuracy through the pre-learned patterns from extensive prior simulation work.
Solution Approach 2:
The patent uses compact models to generate synthetic training data that copies the behavior of full-precision simulation models. This copied data is then used to train deep learning models, allowing the system to learn from extensive simulation results without requiring actual full-simulation runs during the design phase, thereby reducing time while preserving accuracy.
2Manufacturing precision
If traditional simulation-based design methods are used, then device characteristic accuracy can be maintained, but manufacturing cost increases
Solution Approach 1:
The patent employs compact models that are computationally inexpensive and can be rapidly executed to generate large volumes of training data. These compact models serve as disposable, low-cost proxies for expensive full-precision simulations, enabling the training of accurate deep learning models without incurring the high costs of repeated full-simulation runs during manufacturing and design iterations.
Solution Approach 2:
The patent substitutes traditional physics-based simulation mechanisms with data-driven deep learning models. Once trained, the neural networks replace computationally intensive simulation engines, providing comparable accuracy at fraction of the computational cost, thereby reducing manufacturing and design costs while maintaining device characteristic prediction accuracy.
3Measurement precision
If deep learning models are trained with uncertainty analysis, then prediction accuracy improves, but computational complexity increases
Solution Approach 1:
The patent implements uncertainty quantification in the deep learning models to provide feedback on prediction reliability. By analyzing prediction uncertainty, the system can identify cases where the model is less confident and potentially trigger additional verification or data collection, improving overall prediction accuracy while managing complexity through selective application of uncertainty analysis.
Data Source
AI summary
To predict characteristics of a semiconductor device, basic training data corresponding to a combination of process data, device data and simulation result data are generated using a plurality of compact models. Each compact model generates the simulation result data indicating characteristics of a semiconductor device corresponding to the device data by performing simulation based on the device data, the plurality of compact models respectively corresponding to a plurality of process data and a plurality of semiconductor products. A deep learning model is trained based on the basic training data such that the deep learning model outputs prediction data indicating the characteristics of the semiconductor device. Target prediction data indicating characteristics of the semiconductor device included in a target semiconductor product are generated based on the deep learning model, the device data and the process data corresponding to the target semiconductor product.