Neural Network Forward Physical Model Semiconductor Image Reconstruction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for simulating images in semiconductor manufacturing are inefficient due to under-determinacy, noise sensitivity, computational intensity, and the need for precise optical parameters, making them unsuitable for real-time applications and unable to handle phase retrieval effectively.
Innovation Solution
A system that combines a differentiable forward physical model with a neural network to solve inverse problems in semiconductor inspection, metrology, and defect review, using a trainable inversion neural network that incorporates optical, electron beam, and ion beam tools, allowing for real-time processing and improved resolution.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional image simulation methods are used, then image reconstruction can be achieved, but computational complexity is high and processing time is excessive for real-time applications
Solution Approach 1:
The patent pre-trains the neural network model with simulated training data before actual inspection. This preliminary training phase allows the model to learn optimal reconstruction patterns in advance, so that during real-time inspection, the pre-trained model can rapidly process images without requiring extensive computational resources at runtime, thus resolving the contradiction between accuracy and processing time.
Solution Approach 2:
The patent creates a virtual copy of the inspection system through neural network modeling. Instead of performing complex physical computations on actual inspection images, the system uses a trained neural network model that replicates the reconstruction process, enabling real-time processing while maintaining reconstruction accuracy.
2Measurement precision
If traditional inverse problem solving methods are used, then image restoration is possible, but the methods are under-determined and sensitive to noise
Solution Approach 1:
The patent performs preliminary training of the neural network with diverse training data that includes various noise conditions and imaging scenarios. This pre-training establishes robust reconstruction patterns that are less sensitive to noise during actual inspection, resolving the contradiction between restoration quality and noise sensitivity.
Solution Approach 2:
The system uses a feedback mechanism where the neural network's output is compared against expected patterns, and the model parameters are adjusted through backpropagation during training. This feedback loop enables the model to learn optimal reconstruction strategies that minimize noise sensitivity while maintaining high restoration quality.
3Measurement precision
If conventional inspection methods are used, then defect detection can be performed, but resolution is insufficient for small defects on modern semiconductor devices
Solution Approach 1:
The patent creates a virtual high-resolution copy of the defect through neural network super-resolution. Instead of physically upgrading to more complex high-resolution imaging systems, the method uses a trained neural network model to generate high-resolution images from lower-resolution input, achieving improved defect detection resolution while avoiding the complexity of hardware upgrades.
Solution Approach 2:
The patent changes the resolution parameter of the inspection images through neural network processing. By applying super-resolution algorithms, the system transforms low-resolution images into high-resolution images, effectively increasing the detection resolution without changing the physical inspection system, thus resolving the contradiction between resolution and system complexity.
4Measurement precision
If existing image processing algorithms are used, then processing can be performed, but they require precise optical parameters that are difficult to obtain
Solution Approach 1:
The patent enables the neural network model to automatically learn and determine the optimal processing parameters from the training data itself, without requiring external measurement or input of precise optical parameters. The model self-adjusts its parameters during training to achieve accurate reconstruction, resolving the contradiction between processing accuracy and parameter measurement complexity.
Solution Approach 2:
The patent transforms the approach from using fixed, precisely-measured optical parameters to using learnable parameters that are automatically determined by the neural network during training. This parameter transformation allows the system to achieve high processing accuracy without the need for complex external parameter measurement and input processes.
Data Source
Figure 1
Figure 1a
Figure 2
AI summary
Methods and systems for training a neural network are provided. One system includes one or more components executed by one or more computer subsystems. The one or more components include a neural network configured for determining inverted features of input images in a training set for a specimen input to the neural network, a forward physical model configured for reconstructing the input images from the inverted features thereby generating a set of output images corresponding to the input images in the training set, and a residue layer configured for determining differences between the input images in the training set and their corresponding output images in the set. The one or more computer subsystems are configured for altering one or more parameters of the neural network based on the determined differences thereby training the neural network.