Machine learning for polymer degradation prediction : A methodological investigation of validation scenarios, feature representations, and polymer-class effects in automotive thermoplastics

Sammanfattning

Predicting the long-term mechanical degradation of automotive thermoplastics from accelerated ageing data is a challenging problem in industrial materials informatics, especially when random validation does not account for the hierarchical structure of material grades. This thesis investigates how well machine learning can predict elongation‑at‑break retention in polycarbonate/acrylonitrile‑butadiene‑styrene (PC/ABS) and polypropylene (PP) under thermal ageing, and how feature design and validation strategy influence performance. A dataset of 331 specimens from 32 mixed grades provided by Volvo Cars was analysed using six machine-learning algorithms (Lasso, support vector regression, random forest, XGBoost, Gaussian process regression, and a multilayer perceptron) across three scenarios: group‑aware cross-validation, condition‑based forecasting, and temperature‑transfer prediction. The features included polymer class, filler content, ageing descriptors, Arrhenius‑inspired thermal dose, mechanical retention measurements, and FTIR‑derived degradation indices. XGBoost achieved the highest predictive performance across all the validation scenarios. The results showed an R² value of 0.35 for group-aware validation by material grade, whereas the performance was higher for condition-based forecasting (R²=0.69) and temperature‑transfer (R²=0.73). These results demonstrate that the estimated predictive performance is strongly influenced by validation design and increases when material-specific information remains partially represented across the train-test split. Feature selection analysis was done to assess the contribution of different features groups. The results showed that mechanical retention features contribute substantially to predictive performance while FTIR-derived degradation indices did not provide a consistent additional predictive signal both with and without mechanical-retention features. The SHAP analysis identified filler content, mechanical retention features and physics-inspired Arrhenius thermal dose as most influential predictors. Class-specific analysis showed that the predictive performance differs largely between PP and PC/ABS. The larger PP subset supports stronger forecasting and temperature-transfer performance, while the smaller PC/ABS subset shows weaker and less stable performance, mainly in condition-based forecasting. Overall, the work provides a rigorous evaluation framework for machine-learning based prediction in small, heterogeneous and complex polymer‑ageing datasets. The finding shows how validation design, feature representation, polymer class, and data sparsity influence the reliability of ML‑based degradation prediction.