• Home
  • Sitemap
  • Contact us
Article View

Research Paper

Applied Science and Convergence Technology 2023; 32(5): 106-109

Published online September 30, 2023


Copyright © The Korean Vacuum Society.

Machine Learning-Based Prediction of Atomic Layer Control for MoS2 via Reactive Ion Etcher

Changmin Kima , † , Seunghwan Leeb , † , Muyoung Kima , Min Sup Choic , Taesung Kimb , d , and Hyeong-U Kima , *

aDepartment of Plasma Engineering, Korea Institute of Machinery and Materials, Daejeon 34103, Republic of Korea
bSchool of Mechanical Engineering, Sungkyunkwan University, Suwon 16419, Republic of Korea
cDepartment of Materials Science and Engineering, Chungnam National University, Daejeon 34134, Republic of Korea
dSKKU Advanced Institute of Nanotechnology, Sungkyunkwan University, Suwon 16419, Republic of Korea

Correspondence to:guddn418@kimm.re.kr

Received: July 27, 2023; Revised: August 10, 2023; Accepted: August 11, 2023

This is an Open Access article distributed under the terms of the Creative Commons Attribution Non-Commercial License (https://creativecommons.org/licenses/by-nc-nd/4.0/) which permits non-commercial use, distribution and reproduction in any medium without alteration, provided that the original work is properly cited.

This research proposes an innovative method for optimizing plasma etching processes in semiconductor manufacturing using machine learning (ML). Plasma etching is a critical process in defining precise patterns on semiconductor materials, requiring accurate process control. In this study, we employ the ML model based on big data to develop a predictive model that can capture complex relationships between process variables and plasma etching outcomes as the thickness of MoS2. The ML model demonstrated high accuracy, closely aligning with actual experimental results. The experiments confirmed uniform etching across the entire 4-inch wafer, with a precision of approximately 1 nm. Based on this research, we aim to apply ML prediction models to various process conditions of plasma etching and gain deeper insight into the ML’s capabilities for two-dimensional materials in semiconductor manufacturing.

Keywords: Plasma, Machine learning, Layer control, MoS2, Reactive ion etching

In the field of semiconductor manufacturing, the etching processes hold huge importance as a fundamental technique in making intricate patterns on semiconductor material. Etching involves selectively removing specific layers of material through physical or chemical means, playing a crucial role in defining precise geometries and structures necessary for the functionality of integrated circuits and other semiconductor devices [1,2]. In addition, etching enables the creation of nanoscale features essential for realizing advanced semiconductor technologies. As the semiconductor industry continues to advance in miniaturization and device complexity, further developments in etching techniques and process control are expected to drive innovation [3,4].

The prediction and analysis of etching processes through numerical methods and simulations have emerged as indispensable tools for optimizing and understanding the intricate dynamics of this critical fabrication step. Leveraging computational modes, researchers and engineers can simulate the behavior of etchants, the interaction between materials and chemical species, as well as the physical phenomena governing etch rate, selectivity, and surface roughness [5]. These simulations enable systematic investigation of key process parameters, leading to insights into the complex interplay between various factors influencing uniformity and accuracy. Moreover, numerical analysis offers the ability to evaluate the impact of different process conditions on the final device performance, aiding in the design and development of advanced semiconductor devices. As a result, the integration of numerical analysis and simulation techniques into the investigation of etching processes facilitates process control and accelerates technology innovation [6,7].

The persistent trend of increasing semiconductor integration, a reflection of the ever-evolving technology landscape, has heightened the sensitivity of plasma and its interactive reactions, mainly due to the intricate surface microscopic topography. However, this complexity presents several challenges. The plasma reactor, a critical component in semiconductor fabrication, often exhibits abnormal phenomena arising from numerous interactions, including electrical, chemical, optical, and physical processes. The interplay of these factors creates a challenging environment for accurate diagnosis, making it difficult to pinpoint the exact cause behind these anomalies. Consequently, there is an amplified focus on defect control and yield improvement, essential pursuits semiconductor manufacturing. Strategies to mitigate these issues and enhance efficiency are being increasingly implemented, highlighting the importance of consistent evolution and adaptation in this fast-paced field. These ongoing efforts not only ensure the quality and performance of the end products but also advance our capabilities in semiconductor technology, fostering further innovation and growth [8,9].

In the realm of semiconductor manufacturing, accurately predicting wafer thickness is crucial for improving products quality and production process efficiency. Its significance cannot be overstated as it substantially contributes to cost reduction while ensuring product consistency − a critical aspect in maintaining high industry standards. Given these considerations, predictive methodologies leveraging regression analysis are perceived as an optimal approach, providing a robust and reliable framework for addressing this crucial aspect of semiconductor manufacturing [10].

Here, the semiconductor manufacturing relies heavily on etching processes for intricate patterns and selectively remove material layers, defining precise geometries for integrated circuits and devices. Advancements in etching techniques and process control are crucial as the semiconductor industry pursues miniaturization and complexity. Numerical analysis and simulations play a vital role in optimizing etching dynamics, investigating process parameters, and enhancing device design. Defect control and yield improvement are top priorities, driving strategies for enhanced efficiency and adaptation. Accurate wafer thickness prediction is essential for quality and efficiency, with regression analysis as an optimal approach.

In this study, MoS2 films were synthesized on SiO2/Si substrates using an inductively coupled plasma (ICP)-type plasma enhanced chemical vapor deposition (PECVD) system (AFS-IC6T, Allforsystem). The synthesis process involved the deposition of a 1 nm thick Mo metal layer on the SiO2/Si wafer using an e-beam evaporator. The Mo-SiO2/Si wafer was then sulfurized by exposing it to an Ar + H2S plasma in the PECVD chamber. The main plasma parameters used were a radio frequency (RF) power of 550 W, a pressure of 200 mTorr, and a process time of 90 min. Detailed information about the synthesis process can be found in previous studies [1116].

For the plasma-based dry etching process of MoS2 films, an ICPtype reactive ion etching (RIE) system was employed. High-density plasma was generated using an RF generator at 13.56 MHz for the top copper coil. Another RF generator was connected to the bottom side of the electrode to regulate the DC bias voltage. Prior to the etching process, the chamber was evacuated to a vacuum level of 10−6 Torr using a turbomolecular pump (BOC Edwards). The etching of MoS2 films was carried out using an Ar + CF4 + O2 gas mixture, with the gas mixing ratio and chamber conditions precisely controlled to establish optimal etching conditions. Specific details about MoS2 etching can be found in previous studies [7,17].

To determine the number of MoS2 layers, Raman spectroscopy (Alpha 300 M+, WITec GmbH) was performed with 10 single spectra accumulated for each measurement, and an integration time of 1 s per spectrum. The number of layers was further verified using high resolution transmission electron microscopy (HR-TEM) (G2 F30ST, TECNAI, 0.23 nm at 300 kV resolution) with cross-sectional HR-TEM specimens prepared through focused ion beam (FIB) etching (3D FEG, QUANTA, 1.5 nm at 30 kV resolution). The chemical composition of MoS2 was confirmed using X-ray photoelectron spectroscopy (XPS) (AXIS-NOVA, SHIMADZU) with a monochromatic AL-Kα (1486.6 eV) source. Additionally, the atomic force microscopy (AFM) (NX- 10, PARK SYSTEMS) was utilized to confirm the step height of MoS2.

The film thickness and wafer-scale uniformity were measured using a spectroscopic ellipsometer (M-2000V, J. A. Woollam). The detailed sequence of MoS2 synthesis and etching using PECVD and RIE systems is summarized in Table I.

Table 1 . Synthesis and etching sequences of MoS2 by PECVD and RIE.

1Mo deposited onto SiO2/Si wafere-beam evaporator
2Wafers loaded in the ICP PECVD systemPECVD
3Wafers sulfurized by exposure to an Ar/H2S PlasmaPECVD
4Etch chamber cleaning using Ar/O2 plasmaRIE
5Wafers loaded in the load lock chamber using robot armRIE
6Wafers move into the process chamberRIE
7Gas (reactive gas) injectionRIE
8Plasma power on (plasma etching)RIE
9Removing the wafers from the process chamberRIE

This research focused on the predicting wafer thickness under three experimental conditions. There were 9 points of thickness measurement for each of the 76 wafers in our sample set, resulting in a robust dataset comprising of 684 individual data points. The collected data points were divided into training and test sets using ratios of 7.0:3.0, 7.5:2.5, and 8.0:2.0, respectively. By adopting these different division strategies, a comprehensive assessment of the model’s resilience and robustness was aimed to be provided. This rigorous approach to data division ensures not only a balanced perspective in training predictive model but also validates its effectiveness and accuracy across various test scenarios, thereby enhancing the reliability and applicability of findings.

The ICP-RIE system was used for plasma etching that generates plasma by applying power to both the top and bottom of the chamber, as shown in Fig. 1. The experiment focused on etched MoS2, a two-dimensional material whose electrical properties depend on the number of layers. The MoS2 sample was moved to the main chamber through a load lock chamber using robot arm, while maintaining a process pressure of 10−6 Torr and a temperature range of 18–25°C. The detailed process conditions and experiment conditions were presented in Tables II and III.

Table 2 . Process condition of ICP-RIE for MoS2 etching.

Chamber pressuremTorr20
Plasma power (top)W (watt)150, 100, 70, 60, 50
Plasma power (bottom)W (watt)50, 10
Temperature°C20-25 (room temperature)

Table 3 . Experiment conditions of plasma etching for MoS2.

Pressure (mTorr)CF4 (sccm)Ar (sccm)O2 (sccm)Process time (s)Plasma top power (W)Plasma bottom power (W)

Figure 1. ML sequence for prediction of atomic thickness by plasma etching process.

In Table II, we have listed all process conditions for the ICP-RIE used in MoS2 etching, including chamber pressure, plasma power (top and bottom), and temperature. All the conditions were also used for machine learning (ML).

Table III presents the specific experiment conditions for comparison before/after plasma etching of MoS2 with various characterization such as raman, XPS, AFM, HR-TEM, and ellipsometer. Among them, an ellipsometer can measure the total thickness of a wafer over a large area and ascertain uniformity across the entire surface. To apply ML to the data obtained from post-process analysis, first, training data was created and performed pre-processing. The algorithm then generated a ML model to predict the post-processing thickness, enabling a comparison with the actual post-processing thickness. Through iterative development, the model’s accuracy was improved.

The Raman spectrum of the pristine MoS2 exhibited two distinct peaks at 381.12 and 405.71 cm−1, as shown in Fig. 2(a). These peaks were attributed to the in-plane (E12g) and out-of-plane vibrational mode (A1g) of the 2H phase MoS2, respectively [13]. The frequency difference between these E12g and A1g peaks were found to be 24.6 cm−1, which serves as an effective thickness indicator [18]. In response to the reduction in RF top-power, the frequency difference between the E12g and A1g modes exhibited a decreasing trend, measuring 23.6 cm−1 at 150 W, 22.9 cm−1 at 140 W, 22.5 cm−1 at 130 W, and 21.9 cm−1 at 120 W, respectively. These changes in frequency difference correspond to different layers of MoS2. The MoS2 specimen color gradually changed from blue to dark purple correspondingly, as shown in Fig. 2(b).

Figure 2. Characterization of RIE etching process of MoS2. (a) Raman spectrum and optical microscopic image of pristine MoS2 before etching. (b) Raman spectra of plasma etched MoS2 (5-2 L) for RF top-power of 150, 140, 130, 120, 110 W and RF bottom-power of 10 W, respectively. The color of each specimen is given in the inset. (c) XPS spectra of Mo 3d and S 2p for MoS2 before etching. (d) XPS spectra of Mo 3d and S 2p for MoS2 after etching. (e) Cross sectional HR-TEM image of MoS2 before etching. (f) The step height between masked and exposed regions of MoS2 before etching obtained by AFM. (g) Cross sectional HR-TEM image of MoS2 after etching. (h) The step height between masked and exposed regions of MoS2 after etching obtained by AFM. The thickness contour of 4 in. MoS2 is characterized by ellipsometry mapping with 49 points for MoS2 (i) before etching and (j) after etching.

In addition, the XPS measurements [19] were performed to confirm the chemical states in both pristine and etched MoS2, as shown in Figs. 2(c) and 2(d). The examination of the Mo 3d and S 2p provided information about the stoichiometry and crystallinity of the pristine surface [20]. The XPS spectra of the as-grown MoS2 (pristine) displayed two distinct Mo 3d peaks at 232.2 and 229.1 eV, which correspond to Mo 3d3/2and Mo 3d5/2, respectively. Additionally, the S 2p peaks of the pristine sample were observed at 163.1 and 161.9 eV, corresponding to S 2p1/2and S 2p3/2, respectively, as shown in Fig. 2(c). In the case of XPS spectra of the etched MoS2 displayed two distinct Mo 3d peaks at 232.6 and 229.3 eV, which corresponded to Mo 3d3/2and Mo 3d5/2, respectively. In addition, the S 2p peaks of the etched sample were observed at 163.2 and 162.1 eV, corresponding to S 2p1/2and S 2p3/2, respectively, as shown in Fig. 2(d).

To provide clearer evidence of layer-by-layer thinning achieved through RIE, cross-sectional HR-TEM images of the etched sample were prepared using a FIB technique, as shown in Figs. 2(e) and 2(g). Before etching, the four layers of MoS2 can be observed, while three layers remained after etching, demonstrating one layer has been removed by RIE etching. The boundaries between the adjacent layers are distinctly marked using a yellow line for enhanced visibility. The HR-TEM images were in excellent agreement with the raman spectroscopic thickness indicator.

In addition, the step height was measured by AFM, as shown in Figs. 2(f) and 2(h). Before and after etching, the values of step height were 5.02 and 3.64 nm, respectively. The ellipsometer was also used to confirm the overall uniformity of the sample before and after etching. The thickness of the MoS2 on the wafer before etching was measured to be 6.03 nm on average across the entire surface, as shown in Fig. 2(i). The thickness of the MoS2 on the wafer after etching was measured to be 4.76 nm on average across the entire surface, as shown in Fig. 2(j). The ellipsometer measurements showed an overall reduction of about 1.27 nm after the etching process, which was consistent with the TEM results, showing an atomic layer etching of MoS2 in layerby- layer fashion.

In this study, two distinct models were used, namely linear regression and polynomial regression. Figure 3 shows a graph illustrating the relationship between polynomial function degrees and the associated R2 Score in a regression model. As the polynomial degree rises, the R2 Score that indicates model fit also increases (Table IV). The application of polynomial regression is particularly noteworthy due to its capability to capture complex interrelationships among variables. Increasing the degree of the polynomial enhanced the fit between real and predict thickness by R2 Score of models. For a data division ratio of 8.0:2.0, the linear regression model exhibited an R2 Score of 0.743. In contrast, the polynomial regression model displayed an improved R2 Score of 0.897 and 0.954 as the degree increased. Notably, the cubic polynomial regression yielded the highest level of accuracy, highlighting its effectiveness and precision in predicting wafer thickness. The result according to the three experimental conditions is expressed as:

Table 4 . R2 Score results of regression analysis according to degree and data ratio.

Polynomial (degree 2)0.8830.8910.897
Polynomial (degree 3)0.9480.9530.954

Figure 3. Model fit results based on regression analysis methods.


where y is the thickness value, p is the pressure, RfT is top power, and RfB is bottom power. From Eq. (1), we can expect the linear regression form.

It is noted that increasing the degree of polynomial regression inherently escalates the model’s complexity and learning time (Table IV). This consideration becomes critical, especially when dealing with largescale data sets. To overcome these challenges, additional optimization techniques are necessary. Despite potential issues, a perceptible trend was observed as the degree of polynomial regression increased: the prediction results become closer to the y=x line. This indicates an enhancement in the model’s predictive accuracy, providing validity of polynomial regression methods in predicting thickness of MoS2. This emphasizes the potential of polynomial regression as a viable solution for intricate problems in semiconductor manufacturing and beyond.

The plasma etching process is an essential technology in semiconductor manufacturing and product innovation, requiring extensive research. Additionally, investigating various ancillary technologies, such as computer simulation and process diagnostics, is crucial for technological advancement. In this study, we focused on the plasma etching process of MoS2, a two-dimensional material with potential as a nextgeneration semiconductor material. We developed a ML model to predict the post-etching thickness of MoS2. The model’s predictions were found to be in good agreement with the actual experimental results, demonstrating its reliability. Our methodology offers the possibility of creating predictive models with a broader range of experimental conditions, making it applicable to various semiconductor processes with different parameters.

This work was supported by the Technology Innovation Program (or Industrial Strategic Technology Development Program) (20024772) funded By the Ministry of Trade, Industry & Energy (MOTIE, Korea) (1415187508) and KIMM institutional program (NK242F) and NST/KIMM.

  1. H. Abe, M. Yoneda, and N. Fujiwara, Jpn. J. Appl. Phys. 1435, 47 (2008).
  2. K. J. Kanarik, T. Lill, E. A. Hudson, S. Sriraman, S. Tan, J. Marks, V. Vahedi, and R. A. Gottscho, J. Vac. Sci. Technol. A 33, 020802 (2015).
  3. M. Armacost, et al, IBM J. Res. Dev. 43, 39 (1999).
  4. J. Coburn and H. F. Winters, Vac. Sci. Technol. 16, 391 (1979).
  5. S. J. Sherwin, G. E. Karniadakis, and S. A. Orszag, J. Comput. Phys. 110, 373 (1994).
  6. H. K. Kuiken, Proc. R. Soc. Lond. A 392, 199 (1984).
  7. C. Kim, et al, Chem. Mater. 35, 1016 (2023).
  8. R. Clark, K. Tapily, K.-H. Yu, T. Hakamata, S. Consiglio, D. O'Meara, C. Wajda, J. Smith, and G. Leusink, APL Mater. 6, 058203 (2018).
  9. J. Carrere, J. C. Oberlin, and M. Haond, Proceedings of the 2000 5th International Symposium on Plasma Process-Induced Damage (Santa Clara, CA, USA, 2000).
  10. C. Douglas, A. Elizabeth, and G. Geoffrey, Introduction to Linear Regression Analysis. 6th ed. (John Wiley & Sons, 2021).
  11. C. Ahn, et al, Adv. Mater. 27, 5223 (2015).
  12. H. U. Kim, et al, Small 16, 1905000 (2020).
  13. H. U. Kim, et al, ChemSusChem 14, 1344 (2021).
    Pubmed CrossRef
  14. H. Seok, et al, ACS Nano 15, 707 (2021).
    Pubmed CrossRef
  15. H. Seok, et al, ACS Sustainable Chem. Eng. 11, 568 (2023).
  16. H. U. Kim, et al, Appl. Surf. Sci. 470, 129 (2019).
  17. M. H. Jeon, et al, Nanotechnology 26, 355706 (2015).
    Pubmed CrossRef
  18. C. Lee, H. Yan, L. E. Brus, T. F. Heinz, J. Hone, and S. Ryu, ACS Nano 4, 2695 (2010).
    Pubmed CrossRef
  19. M. C. Biesinger, L. W. M. Lau, A. R. Gerson, and R. S. C. Smart, Appl. Surf. Sci. 257, 887 (2010).
  20. F. Lavini, et al, Nanoscale 10, 8304 (2018).
    Pubmed CrossRef

Share this article on :

Stats or metrics