| 226 | 0 | 175 |
| 下载次数 | 被引频次 | 阅读次数 |
碳化硅(SiC)外延层掺杂浓度直接决定功率器件性能，现有调控依赖人工经验与离线测试，存在成本高、滞后性强的问题。为此，提出一种多尺度特征增强时序Transformer(Multi-scale Feature-enhanced Temporal Transformer Network,MFT-Net)模型，整合多尺度卷积、压缩和激励(SE)模块、Transformer及门控循环单元(GRU)模块，构建“当炉–下炉”双场景浓度预测体系。多尺度卷积捕捉毫秒至小时级参数动态，SE强化核心特征，Transformer建模全参数耦合，GRU传递跨炉状态。基于1200炉数据实验表明，模型当炉预测相对误差低至1.35%、决定系数R2达0.89，下炉预测相对误差为1.66%,R2达到0.87，显著优于传统模型。经统计，该模型预计可降低离线测试成本约30%、提升工艺稳定性约15%，为SiC外延智能化提供支撑。
Abstract：Introduction Silicon carbide (SiC),as a representative third-generation wide-bandgap semiconductor material,exhibits superior performance over traditional silicon-based devices in high-voltage,high-power,and high-frequency applications.The doping concentration of the SiC epitaxial layer critically determines the performance of power devices.However,current doping control heavily relies on manual experience and offline testing,which incurs high costs and significant time delays.This study aims to address these limitations by proposing a data-driven model for accurate and real-time prediction of doping concentration in both in-situ and next-batch scenarios.Methods This paper proposes a Multi-scale Feature-enhanced Temporal Transformer Network (MFT-NET) for predicting the core-layer doping concentration during SiC epitaxial growth.The model integrates four key modules:Multi-scale Convolution Module:Employs parallel convolutional branches with kernel sizes of[3,3],[9,3],and[21,5]to capture parameter dynamics across millisecond-to-hour scales.Squeeze-and-Excitation (SE) Module:Enhances feature channels correlated with doping concentration and suppresses noise via channel-wise attention.Transformer Module:Models global coupling relationships among all 75 process parameters using multi-head self-attention.Gated Recurrent Unit (GRU) Module:Captures cross-batch state dependencies to account for residual dopant and equipment state drift between consecutive epitaxial runs.The model was trained and evaluated on a dataset comprising 1200 epitaxial runs from a commercial SiC epitaxial reactor.Data preprocessing included Pearson correlation-based feature selection and min-max normalization.A weighted mean squared error loss function was adopted to prioritize mainstream process samples.Results and Discussion Experimental results demonstrate that MFT-NET significantly outperforms traditional models including Ridge Regression,Support Vector Regression (SVR),CNN,CNN+LSTM,Transformer,and CNN+Transformer.In the in-situ prediction scenario,MFT-NET achieved a mean absolute percentage error (MAPE) of 1.35%,a root mean square error (RMSE) of2.81×1014 cm–3,and a coefficient of determination (R2) of 0.89.For next-batch prediction,it attained a MAPE of 1.66%,RMSE of3.42×1014cm–3,and R2 of 0.87.Ablation studies confirmed the contribution of each module:the multi-scale convolution and SE modules improved feature extraction and selection;the Transformer enabled global parameter interaction modeling;and the GRU was crucial for capturing cross-batch state transfer,reducing next-batch prediction RMSE by 20.8%compared to the model without GRU.Hyperparameter analysis revealed that a 3-layer Transformer,the proposed multi-scale kernel combination,and a Dropout ratio of 0.2 yielded optimal performance.Conclusions The MFT-NET model effectively addresses the multi-scale dynamics,global parameter coupling,and cross-batch state transfer challenges in SiC epitaxial growth.It provides highly accurate doping concentration predictions for both in-situ monitoring and next-batch forecasting,with the potential to reduce offline testing costs by 30%and improve process stability by 15%.This work offers a robust data-driven framework for intelligent optimization of SiC epitaxial processes.
[1]BUFFOLO M, FAVERO D, MARCUZZI A, et al. Review and outlook on GaN and SiC power devices:Industrial state-of-the-art, applications,and perspectives[J]. IEEE Trans Electron Devices, 2024, 71(3):1344–1355.
[2]CHEN X F, YANG X L, XIE X J, et al. Research progress of large size SiC single crystal materials and devices[J]. Light Sci Appl, 2023, 12:28.
[3]GHEZELLOU M, UL-HASSAN J. Influence of different hydrocarbons on chemical vapor deposition growth and surface morphological defects in 4H-SiC epitaxial layers[J]. Phys Status Solidi B:, 2024, 261(4):2300535.
[4]SATO R, KUTSUKAKE K, HARADA S, et al. Machine learning for semiconductor process simulation described by coupled partial differential equations[J]. Adv Theory Simul, 2023, 6(9):2300218.
[5]WANG Z H, LI L, LEON R C C, et al. Improving semiconductor device modeling for electronic design automation by machine learning techniques[J]. IEEE Trans Electron Devices, 2024, 71(1):263–271.
[6]CHEN A, YE S M, WANG Z L, et al. Machine-learning-assisted rational design of 2D doped tellurene for fin field-effect transistor devices[J]. Patterns, 2023, 4(4):100722.
[7]ANGER S, YAYAK A B, ROEDER G, et al. Combining full wafer inspection with deep learning to recognize wafers with critical defects[C]//2023 34th Annual SEMI Advanced Semiconductor Manufacturing Conference(ASMC). Saratoga Springs, NY, USA. IEEE,2023:1–6.
[8]TANG Z C, LI H, LIN P Z, et al. A deep equivariant neural network approach for efficient hybrid density functional calculations[J]. Nat Commun, 2024, 15(1):8815.
[9]KIM T Y, PARK S, LIM C K, et al. Deep learning-based detection of defects in wafer buffer zone during semiconductor packaging process[J].Multiscale Sci Eng, 2024, 6(1):25–32.
[10]WEI S, HAO M L, YU L N, et al. Inverse design of Ga N HEMT heterostructures via deep learning with enhanced auto-encoder framework[J]. Mater Today Commun, 2025, 46:112629.
[11]SALIMIAN A, AMINISHAHSAVARANI A, UPADHYAYA H.Artificial neural networks to predict sheet resistance of indium-doped zinc oxide thin films deposited via plasma deposition[J]. Coatings,2022, 12(2):225.
[12]LIM B, AR1K SÖ, LOEFF N, et al. Temporal fusion transformers for interpretable multi-horizon time series forecasting[J]. Int J Forecast,2021, 37(4):1748–1764.
[13]ZHOU H Y, ZHANG S H, PENG J Q, et al. Informer:Beyond efficient transformer for long sequence time-series forecasting[J]. Proc AAAI Conf Artif Intell, 2021, 35(12):11106–11115.
[14]SHI X J, CHEN Z R, WANG H, et al. Convolutional LSTM network:A machine learning approach for precipitation nowcasting[J]. arXiv preprint arXiv:1506.04214.
[15]DONG R, SONG Y, SIRIWARDANE E M D, et al. Discovery of 2D materials using Transformer network-based generative design[J]. Adv Intell Syst, 2023, 5(12):2300141.
[16]DEY R, SALEM F M. Gate-variants of gated recurrent unit(GRU)neural networks[C]//2017 IEEE 60th International Midwest Symposium on Circuits and Systems(MWSCAS). Boston, MA, USA. IEEE, 2017:1597–1600.
[17]HU J, SHEN L, SUN G. Squeeze-and-excitation networks[C]//2018IEEE/CVF Conference on Computer Vision and Pattern Recognition.Salt Lake City, UT, USA. IEEE, 2018:7132–7141.
[18]BENESTY J, CHEN J D, HUANG Y T, et al. Pearson correlation coefficient[M]//Noise Reduction in Speech Processing. Berlin,Heidelberg:Springer, 2009:1–4.
[19]YU Q Y, LI L, ZHAO H, et al. Evaluation system and correlation analysis for determining the performance of a semiconductor manufacturing system[J]. Complex Syst Model Simul, 2021, 1(3):218–231.
[20]SHAPIRO S S, WILK M B. An analysis of variance test for normality(complete samples)[J]. Biometrika, 1965, 52(3–4):591–611.
[21]WANG Meng, LIANG Yuchao, GU Mei. The application of Pearson's algorithm in semiconductor processes[J]. J China Integrated Circuit,2022, 31(5), 72–75, 81.
[22]LLUGSI R, EL YACOUBI S, FONTAINE A, et al. Comparison between Adam, AdaMax and Adam W optimizers to implement a Weather Forecast based on Neural Networks for the Andean city of Quito[C]//2021 IEEE Fifth Ecuador Technical Chapters Meeting(ETCM). Cuenca, Ecuador. IEEE, 2021:1–6.
[23]TAYMAN J, SWANSON D A. On the validity of MAPE as a measure of population forecast accuracy[J]. Popul Res Policy Rev, 1999, 18(4):299–322.
[24]HODSON T O. Root-mean-square error(RMSE)or mean absolute error(MAE):When to use them or not[J]. Geosci Model Dev, 2022,15(14):5481–5487.
[25]CHICCO D, WARRENS M J, JURMAN G. The coefficient of determination R-squared is more informative than SMAPE, MAE,MAPE, MSE and RMSE in regression analysis evaluation[J]. PeerJ Comput Sci, 2021, 7:e623.
[26]CAMERON A C, WINDMEIJER F A G. An R-squared measure of goodness of fit for some common nonlinear regression models[J]. J Econometrics, 1997, 77(2):329–342.
[27]AWAD M, KHANNA R. Support vector regression[M]//Efficient Learning Machines:Theories, Concepts, and Applications for Engineers and System Designers. Berkeley, CA:Apress, 2015:67–80.
[28]MOSCA, E., SZIGETI, F., TRAGIANNA, S., et al. SHAP-based explanation methods:a review for NLP interpretability[C]//Proceedings of the 29th International Conference on Computational Linguistics.Gyeongju, Republic of Korea. 2022:4593–4603.
基本信息:
DOI：10.14062/j.issn.0454-5648.20250794
中图分类号:TQ163.4
引用信息:
[1]张忠义,王朗,芦伟立,等.基于多尺度特征增强和时序Transformer的SiC外延生长浓度预测模型[J].硅酸盐学报,2026,54(01):35-48.DOI:10.14062/j.issn.0454-5648.20250794.
基金信息:
重点新材料研发及应用国家科技重大专项（2025ZD0619502）
2025-10-31
2025
2025-12-10
2025
1
2026-01-05
2026-01-05
2026-01-05