Corrective Film Pattern Prediction for Semiconductor Wafer Bow
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for reducing wafer bow in semiconductor manufacturing are inefficient due to the long computation times required by finite element methods, and there is a lack of effective machine learning-based solutions to determine corrective film patterns.
Innovation Solution
A machine learning surrogate model, such as a neural network, is used to generate corrective film patterns for reducing wafer bow, trained with a dataset generated from finite element method simulations or experimental data, and optimized using active learning to achieve computation times three orders of magnitude faster than traditional FEM approaches.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If finite element method is used to determine corrective film pattern, then accuracy of wafer bow correction is improved, but computation time increases significantly
Solution Approach 1:
The system performs preliminary finite element method simulations offline to generate a training dataset of wafer bow signatures and corresponding corrective film patterns. This pre-computed knowledge is stored and later used by a trained neural network model that can rapidly predict corrective patterns during actual fabrication without requiring real-time FEM computation.
Solution Approach 2:
The invention creates a surrogate neural network model that copies the predictive capability of the complex finite element method. The neural network learns the mapping between wafer bow signatures and corrective film patterns from FEM-generated training data, then reproduces this knowledge instantly during production, replacing time-consuming FEM simulations with fast neural network inference.
2Productivity
If traditional pixel sum approach is used to determine corrective film pattern, then computation speed is improved, but correction accuracy deteriorates
Solution Approach 1:
The system copies the accurate predictive capability of finite element method into a neural network surrogate model. Instead of using the simplified pixel sum approach, the neural network learns the complex stress distribution patterns from FEM training data, preserving the accuracy of physics-based methods while achieving the speed of machine learning inference.
Solution Approach 2:
The invention transforms the problem from direct FEM computation to neural network inference by changing the computational parameters. The neural network model uses learned weights and biases to directly predict corrective film patterns from wafer bow signatures, bypassing the iterative numerical solving process of FEM while maintaining accuracy through training on comprehensive FEM datasets.
3Adaptability or versatility
If complex device patterns with many thin film materials are used, then device functionality is improved, but wafer bow complexity increases making distortion determination difficult
Solution Approach 1:
The neural network model copies the capability to handle complex multi-material stress interactions by learning from diverse training examples. During offline training, the system exposes the neural network to wafer bow signatures generated from FEM simulations of various thin film material combinations and device patterns, enabling the model to recognize and correct complex bow signatures during production without requiring explicit physics calculations.
Solution Approach 2:
The neural network acts as an intermediary between the complex physical system of multi-material thin films and the corrective film deposition process. Instead of directly calculating stress distributions from first principles for complex material stacks, the neural network learns the effective mapping from observed wafer bow signatures to corrective patterns, simplifying the control of complex physical systems through data-driven modeling.
Data Source
AI summary
A method is disclosed for generating a corrective film pattern for reducing wafer bow in a semiconductor wafer fabrication process. The method inputs to a neural network a wafer bow signature for a predetermined semiconductor fabrication step. The neural network generates from the input a corrective film pattern corresponding to the wafer bow signature. The neural network is trained with a training dataset of wafer shape transformations and corresponding corrective film patterns.