TY - JOUR

T1 - Machine learning pipeline for Structure–Property modeling in Mg-alloys using microstructure and texture descriptors

AU - Guru, Mahish K.

AU - Bohlen, Jan

AU - Aydin, Roland C.

AU - Khalifa, Noomane Ben

N1 - Publisher Copyright: © 2025 The Authors

PY - 2025/8/15

Y1 - 2025/8/15

N2 - Identifying the relationships between material structure and mechanical properties has been crucial for accelerating the exploration of the material design space for advanced alloys. However, traditional approaches for magnesium (Mg) alloys often fall short in providing quantitative and broadly applicable structure–property linkages. To address this challenge, a comprehensive machine learning pipeline is presented for structure–property modeling in extruded Mg-alloys, leveraging both microstructure and texture descriptors derived from experimental data. The pipeline encompasses a robust workflow for data extraction from optical microscopy and X-ray diffraction, advanced image processing and deep learning techniques for microstructure binarization and grain statistics, and the computation of statistical descriptors including n-point spatial correlations, gram matrices for microstructure, and generalized spherical harmonics (GSH) for texture. Dimensionality reduction techniques such as principal component analysis (PCA), isomap, and autoencoders are employed to manage the high-dimensionality of the descriptor space. Subsequently, non-linear regression models—Gaussian Process, XGBoost, and Multi-Layer Perceptron regressors—are evaluated to predict mechanical properties, specifically strain hardening exponent (n) and yield stress (σy). Our results demonstrate that XGBoost consistently outperforms other regressors, achieving a notably low mean absolute percentage error (MAPE) of 6.67% for strain hardening exponent and 7.01% for yield stress, using a combination of PCA-reduced 3-point spatial correlations and isomap-reduced gram matrices as microstructure descriptors, and isomap-reduced GSH coefficients as texture descriptors at a 150μm length scale. Shapley Additive exPlanations (SHAP) analysis further reveals that texture descriptors and aspect ratio distribution are the most influential features in predicting mechanical properties. This established ML framework for structure–property modeling in Mg-alloys, surpasses state-of-the-art benchmarks and provides a valuable template for materials design and discovery.

AB - Identifying the relationships between material structure and mechanical properties has been crucial for accelerating the exploration of the material design space for advanced alloys. However, traditional approaches for magnesium (Mg) alloys often fall short in providing quantitative and broadly applicable structure–property linkages. To address this challenge, a comprehensive machine learning pipeline is presented for structure–property modeling in extruded Mg-alloys, leveraging both microstructure and texture descriptors derived from experimental data. The pipeline encompasses a robust workflow for data extraction from optical microscopy and X-ray diffraction, advanced image processing and deep learning techniques for microstructure binarization and grain statistics, and the computation of statistical descriptors including n-point spatial correlations, gram matrices for microstructure, and generalized spherical harmonics (GSH) for texture. Dimensionality reduction techniques such as principal component analysis (PCA), isomap, and autoencoders are employed to manage the high-dimensionality of the descriptor space. Subsequently, non-linear regression models—Gaussian Process, XGBoost, and Multi-Layer Perceptron regressors—are evaluated to predict mechanical properties, specifically strain hardening exponent (n) and yield stress (σy). Our results demonstrate that XGBoost consistently outperforms other regressors, achieving a notably low mean absolute percentage error (MAPE) of 6.67% for strain hardening exponent and 7.01% for yield stress, using a combination of PCA-reduced 3-point spatial correlations and isomap-reduced gram matrices as microstructure descriptors, and isomap-reduced GSH coefficients as texture descriptors at a 150μm length scale. Shapley Additive exPlanations (SHAP) analysis further reveals that texture descriptors and aspect ratio distribution are the most influential features in predicting mechanical properties. This established ML framework for structure–property modeling in Mg-alloys, surpasses state-of-the-art benchmarks and provides a valuable template for materials design and discovery.

KW - Deep learning

KW - Descriptor

KW - Grain boundary

KW - Machine learning

KW - Magnesium alloys

KW - Microstructure

KW - Orientation distribution function

KW - Property prediction

KW - Structure–Property

KW - Texture

KW - Engineering

UR - http://www.scopus.com/inward/record.url?scp=105006531375&partnerID=8YFLogxK

U2 - 10.1016/j.actamat.2025.121132

DO - 10.1016/j.actamat.2025.121132

M3 - Journal articles

AN - SCOPUS:105006531375

VL - 295

JO - Acta Materialia

JF - Acta Materialia

SN - 1359-6454

M1 - 121132

ER -