A Comparative Analysis of different Machine Learning Algorithms to model Hot Carrier Injection Induced degradation in NMOSFET Planar Transistors

This thesis explores the use of machine learning techniques to model hot-carrier-induced degradation in NMOS planar transistors for 180 nm node and above, using degradation data recorded under controlled electrical stress on gate and source (𝑉𝑔𝑠 and 𝑉𝑑𝑠) of the transistor. Five different machine learning models i.e. Random Forest (RF), Gradient Boosting (XGBoost), Long Short-Term Memory (LSTM), Gaussian Process Regressor (GPR), and Support Vector Regressor (SVR) for degradation modelling (regression modelling) are trained and their performance are compared. The models were tuned and optimized through a systematic tuning called hyperparameter tuning to balance their training efficiency and prediction accuracy. The models were evaluated on training time and the regression evaluation metrices i.e. Root mean square error (RMSE), Mean absolute Error (MAE) and R² score to quantify prediction performance on key degradation indicators such as percentage change in threshold voltage and drain currents in linear and saturated regions. Among the tested models, two machine learning models (Random Forest and XGBoost) demonstrated the highest accuracy and consistency under different stresses (4.5V, 5.4V and 10V 𝑉𝑑𝑠). Further testing explored model’s prediction accuracy dependency on training data volume and the data preprocessing. The findings validate the effectiveness of data-driven modeling for semiconductor reliability prediction and provide a basis for integrating machine learning into Electronic Design Automation (EDA) tools for early-stage reliability assessment.