Predictable etch model using machine learning

Kim, Young Chang; Jung, Sunwook; Kwak, Doohwan; Liubich, Vlad; Fenger, Germain

doi:10.1117/12.2515271

Cited by 6 publications

(7 citation statements)

References 11 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The dimensional geometry features that we use for training is very complex and requires a variety of hidden layers. Therefore, we used a rectified linear unit(=ReLU) to reduce the likelihood of gradient vanishing problem of the hidden layer as an activation function [9] [12]. And we used the ReLU function because we always use positive value such as CD.…”

Section: Ml-rpp Methodologymentioning

confidence: 99%

A new era DFM solution for yield enhancement using machine learning (ML)

Kim,

Kang,

Jung

et al. 2024

DTCO and Computational Patterning III

View full text Add to dashboard Cite

The yield of the deep sub-micron semiconductor is secured by the process capability as well as the yield-friendly design capability. Yield-friendly design capabilities can be equipped with conventional Design for Manufacturability (DFM) that avoids already known defective layouts in design. Previously known defects can be defined as various rules and avoided in design, but defects that may occur at new technology nodes are difficult to avoid in advance. Indiscreetly defect-avoidance designs cause turn TAT increases and Power/Performance/Area (PPA) overheads in the design, which can ultimately lead to increased design costs and poor design competitiveness. The first step of this study is to predict potential risks and to specify major factor of risks that may occur at new process nodes with new DFM solutions developed using Machine Learning (ML) techniques. The second step is to secure early yield through avoidance design to prevent predicted defects and direct mask modification to improve defects. In this study, we present not only the introduction of new ML-based DFM solutions, but also the effect of predicting and improving defects through the application cases of real products.

show abstract

Section: Ml-rpp Methodologymentioning

confidence: 99%

A new era DFM solution for yield enhancement using machine learning (ML)

Kim,

Kang,

Jung

et al. 2024

DTCO and Computational Patterning III

View full text Add to dashboard Cite

show abstract

“…Generally, overfitting becomes worse with more fitting parameters. 2 When overfitting occurs, the neural network model can fit the training dataset well but generalize poorly for the validation dataset.…”

Section: N2e Model Conceptmentioning

confidence: 99%

“…There is a high demand to improve the etch model accuracy for process nodes below 10nm, where the contribution of the etching process to CD uniformity becomes significant. 2 Due to the complexity of the etching process and the interactions among different kernels in a VEB model, the root sources of the residual error that Further author information: (Send correspondence to Zhiheng(Mary) Zuo, E-mail: Mary Zuo@mentor.com, Telephone: 1 503-685-5407)…”

Section: Introductionmentioning

confidence: 99%

A neural network assisted etch model for mask process correction

Zuo

Sharma²,

Bork

2022

Photomask Technology 2022

View full text Add to dashboard Cite

Etch bias correction method is essential to meet the critical dimension (CD) uniformity requirements for mask process correction (MPC), and it has evolved along with the development of process technologies. For matured nodes, rule-based etch bias corrections are adopted. However, this method suffers from limited accuracy and cannot meet the tight CD controls requirement over various patterns for advanced process nodes. To model nontrivial etching process effects such as the aperture effect and the microloading effect, Ref. 1 proposed a variable etch bias (VEB) model. This edge-based semi-empirical model has been widely used in many applications in production and demonstrates good model fits for various layout features and process conditions. In addition, compared to physical etch models, the VEB model is easier to calibrate and requires less runtime. However, for more advanced nodes with EUV masks, and high sensitivity photoresists, only a complex VEB model might be able to meet the precise CD accuracy requirements. The main source of error for the VEB model is the residual error that results from all aspects of the etching process, and a semi-empirical model cannot fully capture it. To overcome this challenge, we propose a neural network assisted etch (N2E) model for MPC. The N2E model is a two-stage etch model that contains a VEB model followed by a neural network assisted model (NNAM). 2 With NNAM, the VEB model in the two-stage N2E model can be simpler than the conventional VEB model while maintaining the same accuracy. In addition, compared to the conventional VEB model, the calibrated N2E model is able to achieve a smaller root mean square error (RMSE) between the measured and predicted etch CDs. Besides, the N2E model produces a small RMSE for the validation dataset and generalizes well. Therefore, the N2E model has the potential to simplify the VEB model part and improve the overall accuracy of MPC.

show abstract

“…The inefficiency of this process, deriving from its redundant steps and significant need for materials, equipment, and staff, starkly contrasts with more streamlined alternatives like ML-based regression modeling. 16,17 This modeling offers efficiency in etch recipe optimization by diminishing the necessary experimental attempts, contrasting with traditional approaches that employ elementary statistical modeling supplemented with deep expertise in the individual process phases. 18 In this research work, our goal was to use a probabilistic ML method, 19 specifically Gaussian process regression (GPR), 20,21 for predicting the etching characteristics.…”

Section: Introductionmentioning

confidence: 99%

High-Accuracy Prediction of ScAlN Thin Film Dry Etching Using Machine Learning Driven Regression Modeling

Shifat,

Jaiswal,

Chityala

et al. 2024

ACS Appl. Electron. Mater.

View full text Add to dashboard Cite

Scandium aluminum nitride (ScAlN) demonstrates notable properties such as high electromechanical coupling, low dielectric permittivity, reduced acoustic losses, and so on. Yet, ScAlN-based device fabrication is challenging, especially controlling the etching process at higher scandium concentrations. This study optimizes the etch recipe by precisely controlling the ScAlN etch rate and sidewall angle using an inductively coupled plasma (ICP) system, characterizing the results with scanning electron microscopy (SEM). We acknowledged that the etch gas flow rates, ICP/RF power, and chamber pressure critically influence the ScAlN etch rate and sidewall angle. Implementing a machine learning (ML) regression model, we predicted these etching outcomes as a function of etch parameters based on historical data correlations, accounting for complex underlying etching physics. We validated the ML model using an independent data set, and it demonstrated over 95% accuracy in predicting ScAlN etch characteristics. This approach significantly enhances the understanding of ScAlN etching, proving ML’s efficiency in etch recipe optimization, which is essential for fabrication of advanced ScAlN-based devices in optoelectronics and RF applications.

show abstract

Predictable etch model using machine learning

Cited by 6 publications

References 11 publications

A new era DFM solution for yield enhancement using machine learning (ML)

A new era DFM solution for yield enhancement using machine learning (ML)

A neural network assisted etch model for mask process correction

High-Accuracy Prediction of ScAlN Thin Film Dry Etching Using Machine Learning Driven Regression Modeling

Contact Info

Product

Resources

About