Lithographic Mask Model Training Using Neural Wafer Pattern Evaluation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The existing lithographic mask generation models face inefficiencies due to complex processing and slow generation speeds when using lithographic physical models to generate wafer patterns, which hampers the training efficiency of these models.

Innovation Solution

A training method utilizing a neural network-based lithographic mask generation model that generates predictive mask maps and wafer patterns, with a pre-trained wafer pattern generation model, and evaluates model precision and mask quality through indices to adjust parameters, reducing computation requirements and improving efficiency.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Loss of time

If a machine learning model is trained using only perfect process window data, then the model training is simple and fast, but the model shows poor generalization ability when encountering actual production data with variations

Engineering Contradiction:
Improvemodel training timeVSAvoidmodel generalization ability
Core Design Contradiction:
Loss of timeVSReliability

Solution Approach 1:

The patent applies preliminary action by pre-processing production data to generate augmented training datasets before model training. Perfect process window data is expanded through multiple transformations (rotation, scaling, translation, noise addition) to create diverse training samples in advance, allowing the model to learn robust features without requiring extensive actual production data during training.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent uses copying by creating synthetic copies of perfect process window data through data augmentation techniques. Multiple transformed versions of the same underlying perfect data are generated to simulate various production conditions, enabling the model to train on abundant synthetic data that mirrors real-world variability without needing to collect and process large amounts of actual imperfect production data.

Inventive Principle:
Principle #26Copying

2Ease of manufacture

If perfect process window data is used for training, then data acquisition is easy, but the data lacks the variations present in actual production environments

Engineering Contradiction:
Improvedata acquisition easeVSAvoiddata representativeness
Core Design Contradiction:
Ease of manufactureVSAdaptability or versatility

Solution Approach 1:

The patent applies parameter changes by systematically transforming the perfect process window data through various parameter modifications including rotation angles, scaling factors, translation offsets, and noise levels. These parameter changes create a diverse set of training samples that reflect actual production variations while maintaining the underlying quality characteristics of perfect data.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent performs preliminary data augmentation to pre-generate diverse training data with various production-like variations. By applying transformations such as adding Gaussian noise, rotating patterns, and scaling features before training, the system prepares a comprehensive dataset that captures production environment variability in advance, making the model adaptable without requiring complex real-world data collection.

Inventive Principle:
Principle #10Preliminary action

3Reliability

If actual production data with variations is used for training, then the model learns real-world variations, but data processing and model training become computationally intensive and time-consuming

Engineering Contradiction:
Improvemodel robustness to variationsVSAvoidcomputational resources required
Core Design Contradiction:
ReliabilityVSPower

Solution Approach 1:

The patent uses copying to create synthetic training data by generating multiple transformed versions of perfect process window data. Instead of processing large volumes of actual imperfect production data, the system creates sufficient training variability through copying and transforming a smaller set of high-quality perfect data, significantly reducing computational resources and training time while maintaining model robustness.

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The patent applies preliminary action by pre-processing and augmenting perfect data to create a comprehensive training dataset before model training begins. This preliminary transformation of perfect data into diverse synthetic samples eliminates the need for computationally intensive processing of actual production data during the training phase, reducing power requirements while achieving the same learning objectives.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4318298B1Training method and apparatus for lithographic-mask generation model, and device and storage medium
Publication Date: 2026.05.20 TENCENT TECHNOLOGY (SHENZHEN) CO LTD
  • EP4318298B1 patent drawingFigure 1~2
  • EP4318298B1 patent drawingFigure 3~4
  • EP4318298B1 patent drawingFigure 5

AI summary

Embodiments of this disclosure provide a training method and apparatus for a lithographic mask generation model, a device and a storage medium, and relate to the field of chip and machine learning technologies. The method includes: generating a predictive mask map corresponding to a chip layout through a lithographic mask generation model (301); generating a predictive wafer pattern corresponding to the predictive mask map through a pre-trained wafer pattern generation model, the wafer pattern generation model being a neural network-based machine learning model (302); determining a model precision evaluation index according to the predictive mask map, and determining a mask quality evaluation index according to the predictive wafer pattern; determining a training loss according to the model precision evaluation index and the mask quality evaluation index (303); and adjusting at least one parameter of the lithographic mask generation model according to the training loss (304). The technical solutions provided in the embodiments of this disclosure improve the training efficiency of the lithographic mask generation model.