Machine learning (ML) has been a widely used tool for recognizing patterns in an extensive database with increasing computing power. This study applied three ML techniques—Cubist, support-vector machine (SVM), and XGBoost (XGB)—to fit the BASELINE surveillance test database. The rule-based Cubist and tree-based XGB models showed a significantly lower root-mean-square deviation than the SVM and ASTM E900-15 nonlinear model. ML showed a good capability for prediction in the interpolation region of the dataset. However, there were significant errors in the extrapolation region in predicting the effect of large fluence on the transition temperature shift (TTS). ML can be very useful in the development of a preliminary model because it can quickly capture the trend of the dataset. To improve the prediction of the TTS trend with fluence in the grouped data with the same initial Charpy impact property, a varying intercept model was introduced into the ASTM E900-15 trend curve, and the model coefficients were estimated by a multilevel modeling procedure. This model considered both the trend of all data and the trend of each group. It then provided more reliable intercepts in each group for TTS prediction with fluence. The multilevel modeling of the grouped datasets can increase the predictive power of the embrittlement trend model for commercial power plants.