Surrogate-assisted optimization of orientation distribution functions for alloy 6061 using physics-based data and explainable machine learning

The crystallographic orientation distribution function (ODF) plays a significant role in controlling the anisotropic mechanical response of aluminum alloy 6061. However, optimization of high-dimensional ODFs still remains a challenge due to the nonlinear nature of microstructure-property relationships and the computational cost of physics-based simulations. This study presents a physics-informed machine-learning framework that identifies optimal ODF configurations to maximize the stiffness–strength synergy metric F Y/E. A synthetic dataset of 150,000 microstructures was generated using Dirichlet-sampled ODFs, rotated stiffness tensors, Voigt-Reuss-Hill homogenization, and a Taylor-type crystal plasticity model. Multi-method feature ranking (Mutual Information, ANOVA, Chi-square, L1-SVC) consistently identified 20 dominant orientation components. A shallow decision tree extracted physically meaningful orientation thresholds that provides transparent processing-relevant rules. A Random Forest surrogate which was trained on scaled ODFs achieved good predictive fidelity (R 2 =0.88). It enabled rapid exploration of the ODF landscape. A multi-start surrogate-assisted optimization strategy was implemented and benchmarked against genetic algorithms and random search. It shows a 2–3% improvement in the best achievable performance. Nearest-neighbor back-projection confirmed that optimized candidates closely matched true physics-based evaluations (<3% deviation). DBSCAN clustering unveiled a unimodal distribution of optimized solutions, which means that only one strong microstructure family existed. In summary, this study opens up a clear and physically based as well as a computationally efficient route for the optimization of microstructures in high-dimensional orientation spaces. Apart from alloy 6061, the suggested framework gives a generalizable basis for the design of anisotropic materials where the mechanical performance is greatly influenced by the crystallographic texture. • Physics-based models generate high-fidelity orientation–property data for alloy 6061. • Multi-method feature ranking identifies the most influential orientation components. • Interpretable machine learning extracts clear orientation rules linked to performance. • Surrogate optimization outperforms genetic and random search in finding optimal texture. • Clustering confirms a single robust high-performance texture family in alloy 6061.