Hybrid Physics-ML Model for Semiconductor Process Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current semiconductor processing lacks a quantitative and objective method to predict the performance of new hardware setups or identify optimal processing parameters, relying on subjective and costly trial-error methods and incomplete physics-based approaches.
Innovation Solution
A hybrid machine learning model is created by combining experimental data from physical substrates with reduced-order physics simulations, using machine learning algorithms to correlate process and hardware parameters, allowing for predictive modeling of on-wafer results for new conditions or tools.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If physics based approaches are used to model semiconductor processes, then theoretical understanding is improved, but the models are incomplete and disparate (separate models for temperature, plasma, and flow)
Solution Approach 1:
The patent combines multiple separate physics-based models (temperature, plasma, flow) into a single integrated hybrid model that simultaneously captures all physical phenomena. This integration is achieved by coupling the physics-based models with machine learning models, creating a unified framework that provides complete process modeling while maintaining the strengths of each individual physics model.
Solution Approach 2:
The patent creates a composite modeling approach by combining two different modeling paradigms: physics-based models and machine learning models. This hybrid model leverages the interpretability and physical consistency of physics-based models while incorporating the predictive power and pattern recognition capabilities of machine learning, resulting in a more robust and comprehensive model.
2Adaptability or versatility
If hardware changes are made to semiconductor processing tools, then process optimization capability is improved, but the cost of implementing hardware changes increases
Solution Approach 1:
The patent performs preliminary virtual experimentation using the hybrid model to predict how process parameters should be adjusted before physically implementing hardware changes. This allows the system to explore different hardware configurations and parameter settings in silico, identifying optimal adjustments without incurring the costs of physical hardware modifications and trial-error experimentation.
Solution Approach 2:
The patent creates a virtual copy of the semiconductor processing tool through the hybrid model, which replicates the tool's behavior and responses to parameter changes. This virtual replica allows for cost-free exploration of hardware modifications and process optimizations, eliminating the need for expensive physical trial-and-error experiments.
3Ease of manufacture
If traditional trial-error methods are used to estimate performance of new hardware, then implementation cost is reduced, but the estimation process becomes subjective and less reliable
Solution Approach 1:
The hybrid model performs self-calibration by automatically adjusting its parameters and predictions based on the data it processes. The machine learning components learn from the physics-based model outputs and experimental data, continuously improving prediction accuracy without requiring subjective human intervention or costly trial-error experimentation.
Data Source
AI summary
Embodiments described herein include processes for generating a hybrid model for modeling processes in semiconductor processing equipment. In a particular embodiment, method of creating a hybrid machine learning model comprises identifying a first set of cases spanning a first range of process and/or hardware parameters, and running experiments in a lab for the first set of cases. The method may further comprise compiling experimental outputs from the experiments, and running physics based simulations for the first set of cases. In an embodiment, the method may further comprise compiling model outputs from the simulations, and correlating the model outputs with the experimental outputs with a machine learning algorithm to provide the hybrid machine learning model.