Mask Pattern Prediction With MRC and Process-Aware Loss
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional machine learning models for generating photolithography masks are not guided by optical proximity correction (OPC) applications and lack awareness of mask rule check (MRC) or process simulation metrics, leading to potential prediction errors and inefficient computing resource consumption.
Innovation Solution
A mechanism is introduced to train a machine learning model using a loss function that considers image reconstruction, MRC evaluation, and process simulation metrics such as edge placement error (EPE) and sub-resolution assist feature (SRAF) printing, ensuring the model minimizes violations and errors in predicting mask patterns.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If conventional machine learning models are used for generating photolithography masks, then the model structure is simple and computing resources are consumed efficiently, but the models are not guided by optical proximity correction (OPC) applications and lack awareness of mask rule check (MRC) or process simulation metrics, leading to prediction errors
Solution Approach 1:
The patent implements a feedback mechanism by incorporating MRC evaluation and process simulation metrics into the loss function during model training. The model receives feedback about its predictions through the loss function, which includes terms for MRC violations and process metrics, allowing the model to iteratively improve its accuracy in generating mask patterns that comply with optical proximity correction requirements.
Solution Approach 2:
The training mechanism is designed to be multi-functional by simultaneously optimizing for multiple objectives: image reconstruction accuracy, MRC compliance, and process simulation metrics. This universal training approach allows a single model to handle various aspects of mask pattern generation, including OPC guidance and MRC awareness, without requiring separate specialized models for each function.
2Reliability
If conventional machine learning models are used without MRC and process awareness, then the training process is faster and less computationally intensive, but the models generate predictions with higher errors and do not minimize MRC violations
Solution Approach 1:
The patent applies preliminary action by performing MRC evaluation and process simulation during the training phase to pre-condition the model on MRC constraints and process metrics. By incorporating these evaluations into the training data and loss function beforehand, the model learns to generate predictions that inherently minimize MRC violations and improve reliability without requiring additional post-processing steps.
Solution Approach 2:
The training process dynamically adjusts parameters including the loss function weights for MRC violations and process metrics based on the specific requirements of each mask pattern generation task. This allows the model to optimize the balance between training time and prediction reliability by emphasizing different aspects of accuracy depending on the specific application requirements.
3Manufacturing precision
If the loss function includes only image reconstruction terms, then the training is simpler and faster, but the model does not minimize MRC violations or process metric errors
Solution Approach 1:
The patent merges multiple evaluation criteria including image reconstruction accuracy, MRC compliance, and process simulation metrics into a unified loss function. By combining these different aspects of quality assessment into a single integrated loss function, the model can simultaneously optimize for all three objectives without requiring separate training processes or complex multi-stage optimization procedures.
Data Source
AI summary
Methods and systems for training a prediction model to predict a mask image in which mask rule check (MRC) violations or process violations (e.g., edge placement error, sub-resolution assist feature (SRAF) printing) are minimized or eliminated. The prediction model is trained based on a loss function that is indicative of (a) a difference between the predicted mask image and a reference image, and (b) at least one selected from: an MRC evaluation of the predicted mask image or an evaluation of a simulated image of the predicted mask image.