The temperature behavior in very high cycle fatigue (VHCF) testing as well as the influence of the intermittent loading is not completely understood. In many cases the high frequency causes the specimens to heat up and may interfere in the material's fatigue performance. In order to address this issue, this study proposed an experimental test with different stress levels and intermittent driving (pulse-pause) with the aid of non-destructive testing (NDT) using a thermography camera. Specimens were coated with black spray to improve the emissivity to 0.93 and conducted to fully reversed condition (R = −1) up to 107 cycles. A large amount of raw data of pulse, pause, stress amplitude, number of cycles and temperature were recorded. These raw data were used to develop tree-based machine learning models called extreme gradient boosting (XGBoost), capable of predicting the temperature throughout the VHCF tests. The result presented a high performance model with determination coefficients (R2) above 0.98, proving the model to be an important ally for ultrasonic fatigue tests. Additionally, Shapley additive explanation (SHAP) method was adopted to assist in the interpretation of the model results.