Machine Learning-Based Prediction of Mechanical Properties of Advanced Engineering Materials: A Review of Data, Models, and Open Challenges

Abstract

Mechanical testing remains the reference method for characterising strength, hardness, stiffness and fatigue resistance, yet it is slow, destructive and impractical for the large compositional and processing spaces now open to advanced engineering materials. Machine learning (ML) models trained on measured or computed data have therefore been proposed as fast surrogates. The relevant literature is spread across material families, and individual studies differ in how inputs are represented, which model is used, and how generalisation error is estimated, so results are hard to compare directly. This article reviews ML-based prediction of mechanical properties for crystalline compounds, steels, high-entropy alloys, concrete and architected composites. The review is narrative and thematic but follows a documented search strategy: representative peer-reviewed studies, databases and benchmarks are organised by data representation, learning algorithm, validation practice and interpretability, and benchmark results obtained under a common protocol are compared quantitatively. The cited work indicates that tree ensembles and neural networks are applied to composition and processing data, that crystal-graph networks become competitive when roughly 10⁴ or more labelled samples are available, and that knowledge-based descriptors can improve on composition alone in alloy design. Recurring open issues include small and heterogeneous experimental datasets, non-uniform validation protocols, data leakage and overfitting, uncertainty quantification, and the limited use of physical knowledge. The article closes with a decision flow for model selection and a reporting-oriented workflow intended to make future studies easier to reproduce and compare. No new experimental results are presented.