Figure 3. Visualization of the theory-based data transformations. Figure 4. Mean surrogate surface using GPR for the original model (a), and for the retrained model (b). (a) (b) data, several combinations were considered and their effect on the model was studied. In GPs, this function determines a significant part of the spatial properties of the model, so its selection is crucial to obtaining an accurate model[13] . The best fit was found to be achieved with a combination of the parameters of a stationary kernel known as the Radial Basis Function (RBF). A first theory-guided ML model was trained using only the single-ramp tests. The resulting mean surrogate surface is shown in Figure 4(a). The train and test scores obtained for this model were 0.9999 and 0.9994, respectively. The trained model was used in a virtual TGA Python module to predict the final yield when varying the temperature cycles. Predictions were performed on both the cycles used for training the model and novel ones not part of the original training data to compare the confidence interval obtained for each. The results are shown in Figure 5, where the purple curve corresponds to real experimental results and the green curve corresponds to the prediction made by the model for the same processing conditions. An iterative process was followed in order www. sampe.org to improve the overall model confidence by conducting new tests and retraining the model with the transformed results. Single-ramp cycle and two-ramp cycle tests were carried out. Figure 4(b) shows a retrain of the original model after analyzing and adding the results from these tests. The train and test scores obtained for the retrained model were 0.9999 and 0.9991, respectively. New predictions on the final yield were performed using the retrained model. Results are shown in Figure 6. A significant improvement in the confidence of the model is observed when comparing the predictions for the same temperature cycle obtained by using the original model against the retrained model, whose achieved confidence values were 69.8% and 99.7% respectively. 3. RESULTS Calculating the conversion rate using a fixed value for the final mass through the analysis of all the studied tests allows us to differentiate the behavior of the material when subjected to various processing conditions. This enables us to obtain accurate predictions of final properties such as the MARCH APRIL 2024 | SAMPE JOURNAL | 45http://www.sampe.org