Small-Signal Modelling of GaN HEMT using ANN, SVR, GPR and Tree-Based Supervised Models Abishkozha Amangeldin, MSc Student School of Engineering and Digital Sciences Electrical and Computer Engineering Nazarbayev University April 26, 2022 Outline Introduction GaN HEMT Objectives Model Development Preprocessing Artificial Neural Network Support Vector Regression Gaussian Process Regression Decision Trees Random Forest ADA Boost Hyperparameter Optimization Loss function Results Discussions Conclusion GaN HEMT High electron saturation velocity Electron mobility Breakdown voltage Operating temperature   Range Step size T(oC) 25-175 25 F(GHz) 0.1-26 0.06475 Objectives To explore ML algorithms for the development of Small-signal modelling To develop ANN, SVR, GPR, DT, RFs and Boosting based models To develop ANN models based on Keras, Scikit-learn and PyTorch libraries To find the optimal architecture of all the developed models To compare the models based on MSE, MAE, R2 and simulation curves Finally, elucidate and discuss the overall comparison Preprocessing Splitting the dataset into training and testing sets Checking for missing data Feature scaling Proposed architecture of the models Artificial Neural Network Support Vector Regression Gaussian Process Regression Decision Trees Random Forest ADA Boost Hyperparameter Optimization Loss function MAE MSE R2 Keras ANN Model Error Estimators, Testing set Testing MSE MAE R2 ReS11 2.6516e-05 0.0029 0.9978 ImS11 6.9419e-03 0.0041 0.9968 ReS21 1.6516e-04 0.0051 0.9945 ImS21 9.1848e-04 0.0046 0.9936 ReS12 2.6785e-05 0.0030 0.9978 ImS12 5.0125e-06 0.0031 0.9942 ReS22 9.5171e-05 0.0041 0.9971 ImS22 4.5291e-05 0.0045 0.9950 ReS11 parameter Scikit-learn ANN Model Error Estimators, Testing set Testing MSE MAE R2 ReS11 4.8561e-03 0.0549 0.9918 ImS11 7.5614e-03 0.0167 0.9904 ReS21 3.8752e-02 0.0579 0.9913 ImS21 5.6846e-03 0.1358 0.9805 ReS12 4.6487e-03 0.0456 0.9935 ImS12 5.9845e-02 0.0845 0.9714 ReS22 2.3516e-03 0.0641 0.9945 ImS22 9.1965e-02 0.1469 0.9913 ImS22 parameter PyTorch ANN Model Error Estimators, Testing set Testing MSE MAE R2 ReS11 6.6841e-03 0.0629 0.9918 ImS11 7.6324e-03 0.0746 0.9904 ReS21 3.8496e-02 0.0483 0.9913 ImS21 4.8217e-03 0.0496 0.9805 ReS12 5.6324e-03 0.0145 0.9935 ImS12 7.4485e-02 0.0674 0.9714 ReS22 8.7436e-03 0.0465 0.9945 ImS22 6.0080e-02 0.0784 0.9913 ImS11 parameter SVR Model Error Estimators, Testing set Testing MSE MAE R2 ReS11 0.0023 0.0015 0.9907 ImS11 0.0168 0.0110 0.9157 ReS21 0.0054 0.0026 0.9548 ImS21 0.0067 0.0031 0.9059 ReS12 0.0130 0.0085 0.9261 ImS12 0.0128 0.0080 0.9345 ReS22 0.0130 0.0199 0.9790 ImS22 0.0258 0.0205 0.9782 ReS22 parameter GPR Model Error Estimators, Testing set Testing MSE MAE R2 ReS11 0.0142 0.0475 0.9867 ImS11 0.0156 0.0674 0.9841 ReS21 0.0247 0.0504 0.9852 ImS21 0.0263 0.0582 0.9782 ReS12 0.0337 0.0699 0.9701 ImS12 0.0316 0.0961 0.9718 ReS22 0.0094 0.0518 0.9910 ImS22 0.0157 0.0989 0.9807 ReS21 parameter DT Model Error Estimators, Testing set Testing set MSE MAE R2 ReS11 0.0071 0.0284 0.9924 ImS11 0.0090 0.0501 0.9909 ReS21 0.0635 0.0572 0.9259 ImS21 0.0319 0.0498 0.9634 ReS12 0.0397 0.0912 0.9547 ImS12 0.0314 0.0954 0.9642 ReS22 0.0191 0.0668 0.9777 ImS22 0.0036 0.0385 0.9923 ImS21 parameter RF Model Error Estimators, Testing set Testing MSE MAE R2 ReS11 0.0102 0.0338 0.9892 ImS11 0.0106 0.0531 0.9891 ReS21 0.0457 0.0417 0.9459 ImS21 0.0199 0.0414 0.9718 ReS12 0.0179 0.0593 0.9940 ImS12 0.0098 0.0503 0.9903 ReS22 0.0082 0.0413 0.9910 ImS22 0.0045 0.0384 0.9948 ReS12 parameter Boosting Model Error Estimators, Testing set Testing MSE MAE R2 ReS11 0.0193 0.0063 0.9823 ImS11 0.0084 0.0155 0.9887 ReS21 0.0062 0.0078 0.9546 ImS21 0.0194 0.0044 0.9936 ReS12 0.0153 0.0055 0.9945 ImS12 0.0087 0.0063 0.9932 ReS22 0.0089 0.0074 0.9912 ImS22 0.0184 0.0181 0.9952 ImS12 parameter ANN Models MSE, Testing set Testing Keras Scikit-Learn PyTorch ReS11 2.6516e-05 4.8561e-03 6.6841e-03 ImS11 6.9419e-03 7.5614e-03 7.6324e-03 ReS12 1.6516e-04 3.8752e-02 3.8496e-02 ImS12 9.1848e-04 5.6846e-03 4.8217e-03 ReS21 5.5785e-05 4.6487e-03 5.6324e-03 ImS21 5.0125e-06 5.9845e-02 7.4485e-02 ReS22 9.5171e-05 2.3516e-03 8.7436e-03 ImS22 4.5291e-05 9.1965e-02 6.0081e-02 Time comparison   DT Ada boost RF PyTorch ANN Scikit-Learn ANN Keras ANN SVR GPR Training Time (min) 0.2 1 2.96 4 7 16 34 72 Models R2, Testing set Testing ANN SVR DT RF Boosting GPR ReS11 0.9978 0.9907 0.9924 0.9892 0.9823 0.9867 ImS11 0.9968 0.9157 0.9909 0.9891 0.9887 0.9841 ReS12 0.9945 0.9548 0.9259 0.9459 0.9546 0.9852 ImS12 0.9936 0.9059 0.9634 0.9718 0.9936 0.9782 ReS21 0.9978 0.9261 0.9547 0.9800 0.9945 0.9701 ImS21 0.9942 0.9345 0.9642 0.9903 0.9932 0.9718 ReS22 0.9971 0.9790 0.9777 0.9910 0.9912 0.9910 ImS22 0.9950 0.9782 0.9923 0.9948 0.9952 0.9807 Conclusions We observed that all the tested models have been able to predict well over a wide range of frequency The DT model is the fastest, followed by the Boosting, RF, PyTorch ANN, Scikit-Learn ANN, Keras ANN, SVR, and GPR models. Keras's ANN model was shown to be the most efficient. The proposed models can be integrated into CAD software for analysis and simulation. Future Works Explore different libraries for other ML models Combination of ML models Bibliography F. Zeng et al., “A Comprehensive Review of Recent Progress on GaN High Electron Mobility Transistors: Devices, Fabrication and Reliability,” Electronics, vol. 7, no. 12, p. 377, Dec. 2018. C. Carpineto, Concept Data Analysis: Theory and Applications, 1st ed. Wiley, 2004. A. Verikas, E. Vaiciukynas, A. Gelzinis, J. Parker and M. Olsson, "Electromyographic Patterns during Golf Swing: Activation Sequence Profiling and Prediction of Shot Effectiveness", Sensors, vol. 16, no. 4, p. 592, 2016. image1.png image2.svg .MsftOfcThm_Accent1_Fill_v2 { fill:#4472C4; } .MsftOfcThm_Accent1_Stroke_v2 { stroke:#4472C4; } image13.png image26.png image27.jpeg image28.png image29.png image30.png image31.png image32.png image33.png image34.png image35.png image36.png image37.png image38.png image39.png image40.png image41.png image42.png