• Machine learning optimizes PEO coatings on 31AZ magnesium alloy. • Support Vector Regression models predict corrosion behavior from process data. • Multi-objective optimization identifies a novel high-performance process window. • Optimized coating shows dramatic improvement in corrosion resistance. • Prospective validation confirms model accuracy and process robustness. Plasma Electrolytic Oxidation (PEO) is a highly effective surface treatment for enhancing the corrosion resistance of lightweight magnesium alloys, but its performance is critically dependent on a complex interplay of process parameters. Traditional optimization methods are often resource-intensive and struggle to capture non-linear process–property relationships under experimentally conditions. In this study, we demonstrate a local, proof-of-concept machine learning-guided optimization framework to systematically explore a restricted PEO process window for enhanced corrosion protection of AZ31 magnesium alloy. Support Vector Regression (SVR) was employed to build surrogate models mapping current density and treatment time to corrosion potential (E corr ) and corrosion current density (I corr ). A comparative benchmark against Artificial Neural Networks (ANNs) confirmed numerical stability of SVR in this small-data regime. The SVR-based Icorr model demonstrated statistically meaningful predictive capability (R² = 0.58), while the Ecorr model was retained as a qualitative directional constraint rather than a quantitatively predictive surrogate. These surrogates were coupled with a Differential Evolution algorithm to solve a constrained multi-objective local optimization problem. The framework identified a previously untested optimal condition of 2309 mA/cm² and 218 s within the experimentally explored domain. A prospective validation experiment confirmed the practical effectiveness of the optimization outcome, yielding a coating with an exceptionally low mean corrosion current density of 0.65 ± 0.26 µA/cm²—representing a substantial improvement relative to all initial experimental conditions. Despite the limited dataset, this work demonstrates that carefully constrained surrogate-based optimization, when coupled with prospective validation, can provide actionable guidance for process improvement in data-scarce PEO systems.
Mohammadipour et al. (Wed,) studied this question.