Quantifying Algorithmic Decision-Making in Engineering: A High-Fidelity Explainable AI (XAI) Framework for Non-Linear Regression

Abstract

Applying machine learning to safety-critical materials science is often limited by the blackbox nature of AI models. This study presents an explainable AI (XAI) framework to balance predictive accuracy with physical interpretability. There are four distinct architectures (Linear Regression, Support Vector Regression (SVR), Gradient Boosting, and Random Forest) in this study, using a Test Split R² and 10-fold Cross-Validation to ensure results are generalisable. SHapley Additive exPlanations (SHAP) method is used to study the effect of different factors that can affect materials performance. This approach follows Green AI concept, without the requirements of resource-intensive laboratory sessions, it provides the same prediction values.

This study successfully identified the significant effect of fracture energy under low temperature testing, where it showed a sharp sensitivity threshold at -40 °C. The use of explainable AI (XAI) helps to provide more insights and practical predictions for further engineered designs and enhancements. It also reduced the use of non-recyclable materials required in laboratory. This study showed that by using explainable AI models, even with limited experimental results can provide additional insights and practical predictions for further engineered designs and enhancements.