FEATURE / HIGH-TEMPERATURE COMPOSITES (a) (b) Figure 5. Comparison of final yield predictions made using the original model for a test used for the training (a) and a new test (b). (a) (b) Figure 6. Comparison of final yield predictions made using the original (a) and retrained (b) models for the 1ÂșC/min single-ramp temperature cycle. final yield for specific processing conditions and compare them to other potential options. Figure 5(a) shows both the experimental results and the predicted final yield of a TGA test. A confidence of 97.1% was achieved by the model for this prediction. Figure 5(b) shows the prediction made for a potential test with lower confidence of 69.8%. As previously mentioned, the confidence achieved when predicting the same temperature cycle was improved by 29.9% just by adding one additional test to retrain the ML model; and the final yield prediction for an unknown two-ramp cycle was also performed with 86.7% confidence. Results obtained for the latter are shown in Figure 7. Figure 4 shows the evolution of the mean surrogate surface when retraining the model with data from more complex temperature cycles. The difficulty in fitting temperature cycles with varying heating rates can be observed in Figure 4(b). Nonetheless, results obtained for this model, including the train and test scores, as well as the confidence of the performed predictions suggest that the selected combination of covariance function and theory-based dimensions succeeds in characterizing the process and predicting final properties for more complex processing conditions as well. 46 | SAMPE JOURNAL | MARCH APRIL 2024 4. SUMMARY AND CONCLUSIONS While existing theories can potentially characterize the pyrolytic process of polymeric systems subjected to simple heating cycles, these models are specific to certain processing conditions and cannot accurately predict material behavior when these conditions are not met. This represents a challenge for the industry seeking optimal manufacturing parameters for the desired outcome. By analyzing the conversion rate of pyrolysis, it is possible to predict the final yield for different processing conditions by training theory-guided machine learning models such as GPs. The proposed framework in this study has been successful in predicting the final yield of a material subjected to high-temperature thermal cycles with relatively little experimental data. The iterative process of retraining the model has been effective in significantly improving the accuracy and confidence in the prediction of complex processing cycles. This method can therefore be implemented to accelerate the characterization and analysis of the pyrolytic process in the manufacturing of composite materials for high-temperature applications with reasonable accuracy while minimizing experimental efforts and resources. www. sampe.orghttp://www.sampe.org