ML Image Alignment for Lithography Pattern Reproduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing lithographic processes face challenges in accurately reproducing patterns with dimensions smaller than the classical resolution limit, necessitating sophisticated techniques like optical proximity correction and customized illumination schemes to overcome difficulties in achieving desired electrical functionality and performance.
Innovation Solution
A machine learning model is trained to generate predicted measured images by aligning non-aligned input target and reference measured images using a generator-discriminator framework, iteratively adjusting parameters to minimize discrepancies and align images based on specified thresholds.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If traditional lithographic processes are used to reproduce patterns with dimensions smaller than the classical resolution limit, then manufacturing precision deteriorates, but using sophisticated techniques like optical proximity correction and customized illumination schemes increases device complexity
Solution Approach 1:
The patent uses a generative adversarial network (GAN) to create a computational model that copies the mapping relationship between target images and measured images. The GAN learns from training data to generate predicted measured images that accurately represent the lithographic process outcomes, replacing complex physical correction techniques with a data-driven computational approach.
Solution Approach 2:
The patent transforms the lithographic process modeling from physical parameter adjustments (optical proximity correction, illumination schemes) to statistical parameter learning through machine training. The GAN models the complex relationships between input and output images by learning from training data, changing the approach from physics-based parameter tuning to data-based parameter optimization.
2Measurement precision
If image alignment is performed using traditional methods, then measurement precision is insufficient for sub-resolution features, but the patent's machine learning approach requires additional training and computational resources
Solution Approach 1:
The patent replaces traditional mechanical and optical image alignment methods with a machine learning-based computational approach. The GAN-based system substitutes physical alignment mechanisms with algorithmic image generation and comparison, enabling sub-resolution feature alignment through computational intelligence rather than mechanical adjustment.
Solution Approach 2:
The GAN model performs self-alignment by generating predicted measured images that inherently contain the alignment information. The system automatically learns the alignment transformations during training and applies them during inference, eliminating the need for separate manual or mechanical alignment procedures.
Data Source
AI summary
A method for training a machine learning model to generate a predicted measured image, the method including obtaining (a) an input target image associated with a reference design pattern, and (b) a reference measured image associated with a specified design pattern printed on a substrate, wherein the input target image and the reference measured image are non-aligned images; and training, by a hardware computer system and using the input target image, the machine learning model to generate a predicted measured image.