Semiconductor Chamber Digital Twin for Fast Process Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current semiconductor manufacturing processes face challenges in predicting the performance of deposition, etch, and clean processes due to the high computational cost and time required by physics-based simulations, and the limitations of data-driven models in capturing complex plasma environments, leading to lengthy design and testing cycles.
Innovation Solution
Hybrid models combining physics-based and data-driven approaches, utilizing machine-learning techniques, are employed to integrate disparate datasets from simulations and experiments, creating a predictive and efficient system model that accelerates design and optimization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If physics-based simulations are used to predict process performance, then prediction accuracy is improved, but computational cost and time increase significantly
Solution Approach 1:
The patent creates a digital twin (virtual copy) of the semiconductor manufacturing tool that replicates the physical system's behavior. This digital twin is trained using physics-based simulations and experimental data, allowing it to predict process performance with high accuracy while requiring minimal computational resources during operation, thus resolving the contradiction between prediction accuracy and computational time
Solution Approach 2:
The patent performs comprehensive physics-based simulations and data collection in advance to train the digital twin model before actual manufacturing operations. By pre-computing the complex physics-based models and storing the results in a trained digital twin, the system eliminates the need for real-time physics-based simulations during production, significantly reducing computational time while maintaining prediction accuracy
2Loss of time
If data-driven models are used to predict process performance, then computational time is reduced, but ability to capture complex plasma environments deteriorates
Solution Approach 1:
The patent merges physics-based models with data-driven machine learning models to create a hybrid digital twin. The physics-based component captures the fundamental plasma physics and chemical reactions, while the machine learning component learns from experimental data to correct and enhance predictions. This combination maintains the ability to capture complex plasma environments while achieving fast computational speeds suitable for real-time applications
Solution Approach 2:
The digital twin employs a composite modeling approach, integrating multiple modeling paradigms (physics-based simulations, reduced-order models, and machine learning models) into a unified predictive system. This composite structure leverages the strengths of each individual model type: physics-based models provide fundamental accuracy, reduced-order models provide computational efficiency, and machine learning models provide adaptability to experimental data, collectively achieving both speed and accuracy
3Loss of information
If high-fidelity physics-based simulations are used, then understanding of physical mechanisms is improved, but design and testing cycles lengthen
Solution Approach 1:
The patent segments the digital twin into multiple independent but interconnected models operating at different spatial and temporal scales: chamber-scale physics models, feature-scale plasma models, and wafer-scale process models. This segmentation allows each model to be optimized for its specific purpose and trained independently, enabling parallel development and faster integration, thus accelerating design cycles while maintaining comprehensive physical mechanism understanding
Solution Approach 2:
The digital twin is designed as a dynamic, adaptive system that can adjust its modeling fidelity and computational resources based on the specific prediction task. For routine predictions, the system uses pre-trained, computationally efficient models. When novel process conditions or new physical phenomena need to be explored, the system can dynamically invoke higher-fidelity physics-based simulations, optimizing the balance between understanding physical mechanisms and maintaining fast design cycles
Data Source
AI summary
A computer-implemented method builds and uses a reduced-order model of a semiconductor manufacturing chamber. The method generates a design-of-experiments matrix that defines combinations of chamber parameters. It executes physics-based simulations to obtain training data for the combinations. The training data includes physical quantities and derived values. The method trains a machine-learning algorithm using the training data to create the reduced-order model. The reduced order model maps the chamber parameters to predicted performance metrics. The method receives a chamber configuration as input to the reduced-order model. It generates predicted performance metrics for the configuration. The method controls at least one process recipe of the chamber based on the predicted metrics. The method presents results on a display. The method stores an updated process recipe definition in a hardware-based memory. The method transmits the updated process recipe definition to a controller of the chamber for execution.