Inverse Lithography Mask Optimization via Machine Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current inverse lithography technologies rely heavily on numerical solvers, which are slow and dependent on initial conditions, leading to sub-optimal results, and require significant intervention, limiting the effectiveness of mask optimization in semiconductor manufacturing.
Innovation Solution
A method utilizing a machine learning model to predict an optimized mask image by processing input mask and design images during the inverse lithography process, allowing for iterative refinement of the mask to achieve a minimized error in pattern transfer onto semiconductor wafers, leveraging implicit layers and unrolled neural networks for efficient training and deployment.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If numerical solvers are used for mask optimization in inverse lithography, then the process can be performed with traditional methods, but the computation speed is slow and the results are sub-optimal due to dependency on initial conditions
Solution Approach 1:
The patent replaces the traditional numerical solver-based iterative optimization process with a machine learning model (neural network) that has been trained to directly predict optimized mask patterns. This substitution eliminates the need for repeated forward and backward computation passes through numerical solvers, thereby dramatically improving computation speed while maintaining or enhancing optimization result quality through the model's ability to learn optimal solutions from training data
Solution Approach 2:
The patent performs preliminary training of the machine learning model using numerical solvers during an offline phase to create a pre-trained model that encapsulates optimization knowledge. During actual mask optimization, this pre-trained model is deployed to rapidly predict optimized masks without requiring real-time numerical solver execution, thus achieving fast computation speeds while preserving the optimization capabilities developed during the preliminary training phase
2Reliability
If machine learning is used to compute the initial mask, then the initial condition is improved, but the solution still relies on numerical solvers for forward/backward computation
Solution Approach 1:
The patent merges the functions of initial mask computation and iterative optimization into a single integrated machine learning model. The neural network is trained to perform both tasks simultaneously, eliminating the need for separate numerical solver steps for forward and backward computation. This consolidation reduces process complexity while maintaining the benefits of improved initial mask quality
3Manufacturing precision
If iterative forward and backward computation is performed, then the mask can be optimized, but the process is time-consuming and highly dependent on initial conditions
Solution Approach 1:
The patent substitutes the time-consuming iterative forward and backward computation process with a pre-trained machine learning model that directly predicts optimized mask patterns. The model was trained during an offline phase using numerical solvers to achieve high pattern transfer fidelity, and during deployment, it rapidly generates optimized masks without requiring real-time iterative computation, thus dramatically reducing optimization time while maintaining manufacturing precision
Data Source
AI summary
In the semiconductor industry, lithography refers to a manufacturing process in which light is projected through a geometric design on a mask to illuminate the design on a semiconductor wafer. The wafer has a light-sensitive material (i.e. resist) on its surface which, when illuminated by the light, causes the design to be etched onto the wafer. However, this lithography process does not perfectly transfer the design to the wafer, particularly because some diffracted light will inevitably distort the pattern etched onto the wafer (i.e. the resist image). To address this issue in lithography, an inverse lithography technology has been developed which optimizes the mask to match the desired shapes on the wafer. The present disclosure improves current inverse lithography technology by employing machine learning for mask optimization.