Machine Learning Model for Predicting Process-Sensitive Geometries
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional optical proximity correction (OPC) techniques, such as optical rule checking (ORC) simulations, face challenges in accurately identifying process-sensitive geometries (PSGs) in integrated circuits (ICs), leading to potential manufacturing defects due to under or overcompensation for manufacturing variations, especially at nanometer scales.
Innovation Solution
A method using a predictive model based on a training data repository of ORC simulations to predict PSGs in IC layouts, with iterative adjustments based on actual manufacturing data to improve prediction accuracy, flagging correct or incorrect predictions and updating the model accordingly.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional ORC simulations are used to identify PSGs, then manufacturing sensitivity can be detected, but prediction accuracy deteriorates due to under or overcompensation
Solution Approach 1:
The patent implements a feedback mechanism where actual manufacturing data from produced ICs is fed back to retrain and refine the predictive model. The system compares predicted PSGs with actual PSGs identified from manufactured circuits, uses the discrepancy to adjust model parameters, and iteratively improves prediction accuracy. This closed-loop feedback resolves the contradiction by continuously enhancing measurement precision while reducing manufacturing defect risk through learned corrections.
Solution Approach 2:
The patent changes the parameters of the predictive model based on actual manufacturing data. By adjusting model parameters through retraining with real-world manufacturing outcomes, the system transforms the fixed-parameter conventional ORC simulation into an adaptive model that accurately reflects actual manufacturing processes, thereby improving both PSG identification accuracy and reliability.
2Manufacturing precision
If manual or conventional OPC techniques are applied to increase chip resolution, then manufacturing precision improves, but device complexity increases
Solution Approach 1:
The patent replaces manual or conventional mechanical/optical OPC adjustment processes with an automated machine learning-based predictive model. Instead of relying on complex manual interventions or conventional iterative OPC simulations, the system uses trained ML models to predict PSGs and guide corrections, substituting automated intelligent processing for complex manual or conventional procedures, thereby maintaining manufacturing precision while reducing operational complexity.
3Measurement precision
If ORC simulations are applied under multiple process variants to identify sensitive geometries, then manufacturing sensitivity detection improves, but loss of time increases due to extensive simulations
Solution Approach 1:
The patent performs preliminary actions by pre-training the predictive model on extensive ORC simulation data from multiple process variants before actual manufacturing. The model learns the relationships between process variations and PSG formation in advance, so during actual production, predictions can be made quickly without running extensive simulations each time. This resolves the contradiction by investing time upfront in model training rather than continuously during manufacturing.
Solution Approach 2:
The patent creates a virtual copy of the manufacturing process through the trained predictive model. Instead of repeatedly running time-consuming physical or detailed simulations, the system uses the trained model to replicate and predict outcomes based on learned patterns from previous simulations and actual data, dramatically reducing the time required for sensitivity detection while maintaining accuracy.
Data Source
AI summary
Methods according to the disclosure include: predicting process-sensitive geometries (PSGs) in a proposed IC layout based on violations of a set of processing constraints for the proposed IC layout, the set of processing constraints being calculated with a predictive model based on a training data repository having a plurality of optical rule check (ORC) simulations for different IC layouts; identifying actual PSGs in a circuit manufactured using the proposed IC layout; determining whether the predicted PSGs correspond to the actual PSGs in the manufactured circuit as being correct; in response to the predicting being incorrect: adjusting the predictive model based on the actual PSGs, wherein the adjusting includes submitting additional ORC data to the training data repository; and flagging the proposed IC layout as incorrectly predicted; and in response to the predicting being correct, flagging the proposed IC layout as correctly predicted.