Uncertainty-aware active-learning design of high-strength, corrosion-resistant as-extruded Mg–Al–Zn–Mn–Ca–Y alloys using Bayesian neural networks

Abstract

The present study proposes an uncertainty-aware design framework for as-extruded Mg–Al–Zn–Mn–Ca–Y alloys by integrating Bayesian neural network (BNN)-based property prediction, uncertainty-driven active learning (AL), and experimental validation. In the present AL cycle, tensile-property predictive uncertainty was used to identify informative composition–extrusion-temperature conditions with high uncertainty, whereas corrosion rate was measured for specimens selected from the AL-identified uncertainty-rich region. The framework was applied to explore the strength–corrosion trade-off in Mg alloys using in-house experimental data obtained under controlled manufacturing conditions. Experimental validation showed that the AL-derived alloys exhibited favorable strength–corrosion combinations relative to the initial dataset. Among the validation specimens, the Mg–9.57Al–1.47Zn–0.45Mn–0.72Ca–0.89Y (wt%) alloy extruded at 400 °C showed the most favorable balance, with a tensile yield strength of 234 MPa, ultimate tensile strength of 343 MPa, elongation of 9.8%, and corrosion rate of 0.66 mm/y. This favorable balance was consistent with stronger basal texture, equilibrium-predicted solute/particle-partitioning trends, and electrochemical observations—relatively low cathodic current densities and an impedance response fitted with a film-resistance element—that were consistent with the lower corrosion rate determined by mass-loss testing. The results demonstrate that predictive uncertainty can guide metallurgically interpretable experiments for alloy design under limited data conditions. The proposed BNN/AL framework should be regarded as an uncertainty-guided exploration strategy rather than a deterministic property optimizer.