This article discusses the relationship between machine learning (ML) and semiconductor manufacturing, specifically the application of ML algorithms and models in the semiconductor manufacturing process.
Image Credit: Gorodenkoff/Shutterstock.com
Importance of ML in Semiconductor Manufacturing
In recent years, significant advancements in ML have influenced several fields beyond computer science, including autonomous driving, structural color design, medicine, and face recognition.
The capabilities of ML can be utilized to optimize and automate the semiconductor manufacturing process and the associated data analysis. Several studies have been performed to develop and apply different ML algorithms and models in semiconductor manufacturing, including malfunction detection, device production, process optimization, and wafer detection.
Supervised discriminative ML models that typically learn from experience can improve future semiconductor design and manufacturing efficiency owing to the sufficient availability of well-labeled historical data obtained from the existing semiconductor manufacturing process.
For instance, a multilayer perceptron (MLP) classifier model can identify inking patterns from the failure map automatically and perform postprocessing for correction to eliminate the need for human intervention in die inking during die screening.
Similarly, supervised generative ML models are typically used to replace the manual design for improved manufacturability. For instance, conditional generative adversarial network (CGAN) based WellGAN can be used in the layout design phase to generate well layouts for analog and mixed-signal (AMS) circuits automatically, replacing manual design.
CGAN can also be applied in manufacturing processes such as lithography to effectively model three-dimensional (3D) aerial images and resist patterns based on the given mask patterns to increase manufacturing efficiency significantly.
Applications of ML in Semiconductor Manufacturing
Yield Prediction and Analysis
Predicting the product yield and understanding the impact of process parameters on the yield is crucial for semiconductor manufacturing, as the decline in product yield and changes in process parameters are correlated with each other. Supervised discriminative ML algorithms such as regression and convolutional neural networks (CNN) for pattern recognition can automatically identify such correlations.
For instance, multivariate adaptive regression splines (MARS) with genetic algorithm (GA)-selected features can be applied to effectively estimate the production yield across various design generations and fabrication processes, leading to a significant reduction in the number of required fabrication or simulation data.
Detection of Manufacturing Process Excursions
Early detection of process extrusions is crucial to avoid substantial scrap and test costs and potential quality issues. Unsupervised generative ML algorithms such as GAN or autoencoder for hot spot checking and abnormal pattern recognition can be employed for excursion detection based on probe wafer maps and process tool data.
Simplification of Manufacturing Flow
Unsupervised ML algorithms can also be utilized to optimize the manufacturing test flow, such as enabling package burn-in (BI) elimination, to identify “risky” materials and send them to BI stress. For instance, the kernel-based clustering (KBC) algorithm, a type of unsupervised learning, can identify the potential cluster defect based on wafer probe test data and send the risky dies to package BI.
Improvement in Design for Manufacturability (DFM) Tools
ML, specifically deep neural networks (DNN), can be used to improve and automate the DFM tools and checkers considering all key metrics and critical inputs. A DNN can be trained to predict potential design failures/violations.
The inputs can include previous customer quality complaint (CQC) database info, yield criteria, technology metal options, physical integration, and design rule checks (DRCs). Scoring guidelines for every requirement and a pass/fail criterion with corresponding optimization goals can also be used as inputs.
The decision-making of the neural networks can be interpreted to determine the critical design/layout features that are correlated with the predicted failures/violations. Critical features can be identified based on the high sensitivities of the output decision and provided as feedback to the test/design process.
Other ML Applications
Differential evolution (DE) algorithm and DNN can be used to increase the wafer productivity to improve the return on investment (RoI) by considering the selling and cost prices of the wafer as optimization factors and RoI as the optimization goal.
An automatic defect classification (ADC) system using a scanning electron microscope image as input can be employed to classify and identify wafer surface defects. The system can effectively perform detection without human intervention using the CNN model.
Similarly, a transfer learning method based on CNN can be utilized for wafer defect classification to significantly reduce the cost of ML computation. Studies demonstrated that the method could classify defect images with greater accuracy.
Inaccurate wafer pass/fail tests can adversely affect the entire semiconductor manufacturing. ML can be employed to predict and reduce process malfunctions. For instance, a deep belief network (DBN)-based multi-classifier can automatically assess the wafer test by collecting signals from sensors in processes and then effectively predicting the fault detection.
Conclusion and Future Outlook
To summarize, ML algorithms can be used effectively in several areas of semiconductor manufacturing for automation and optimization. However, more research is required to address several existing challenges regarding the application of ML.
For instance, the training data required to build effective general ML models are often insufficient. The degree of generalization for the ML goal must match the size of the training data. Additionally, all examples in the training data must consistently represent the hidden relation between the target and the input.
More from AZoM: Can Blockchain Technology Better Plan Semiconductor Supply Chains?
References and Further Reading
Liu, D., Xu, L., Lin, X., Wei, X., Yu, W., Wang, Y., Wei, Z. (2022). Machine learning for semiconductors. Chip, 1(4), 100033. https://doi.org/10.1016/j.chip.2022.100033
C. He., H. Hu., P. Li. (2021). Applications for Machine Learning in Semiconductor Manufacturing and Test (Invited Paper). 2021 5th IEEE Electron Devices Technology & Manufacturing Conference (EDTM). https://web.ece.ucsb.edu/~lip/publications/MLSemiManufacturingTestIEEE-EDTM2021.pd
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.