Machine Learning in Plasma Electrolytic Processes: Prediction Models, Data-Driven Optimization, and Process Digitalization - A Review

Plasma electrolytic oxidation (PEO) is a plasma-assisted anodic treatment for light alloys that produces hard, adherent ceramic-like oxide layers with improved corrosion and wear performance [۱,۲,۴,۵]. The process is, however, highly nonlinear and controlled by coupled electrical and electrochemical phenomena that depend sensitively on power-supply waveforms, electrolyte chemistry and alloy composition [۱,۲,۵,۶]. Consequently, the design of coatings with targeted properties still relies largely on empirical trial-and-error approaches, and the systematic use of machine learning (ML) and artificial intelligence (AI) for process digitalization is only beginning to emerge [۳,۷,۸]. At the same time, ML/AI methods have become powerful tools in materials science and corrosion engineering, where they are used to learn structure-property relationships and accelerate optimisation [۹-۱۱]. In recent years several groups have started to transfer these concepts to PEO and related micro-arc / plasma-electrolytic oxidation processes. Reported applications include prediction of coating thickness and roughness, electrochemical and tribological performance, morphology classes and optimal process windows [۳,۷,۱۲-۱۶]. However, these contributions are scattered across different journals and alloy systems, and no focused overview exists of which algorithms, descriptors and target variables are used, how models are validated, and which best practices are emerging. This work therefore provides one of the first dedicated reviews of ML/AI applications in PEO and plasma-electrolytic coating technologies. A systematic literature search was carried out in Scopus, Web of Science and Google Scholar up to November ۲۰۲۵, combining "plasma electrolytic oxidation", "micro-arc oxidation" and "electrolytic plasma" with "machine learning", "artificial intelligence", "deep learning" and related terms. Studies were retained when a PEO or closely related plasma-electrolytic process was used to produce a coating or modified surface, and an ML/AI model was trained on experimental or simulation-derived data to predict, classify or optimise coating morphology, properties, electrochemical behaviour or process parameters. For each paper, we recorded the material system, processing conditions, data origin and size, ML algorithm family and key modelling choices and performance metrics. Across the surveyed literature, three main application domains emerge. The first concerns prediction and optimisation of coating thickness, morphology and energy efficiency in PEO of aluminium and titanium alloys, where tree-based ensembles and gradient-boosting models, often coupled with Shapley additive explanations (SHAP), rank the influence of pulse parameters, current density and electrolyte composition on coating growth [۳,۱۲]. A second domain targets the corrosion and electrochemical response of PEO- or MAO-coated biodegradable magnesium alloys, using artificial neural networks, Gaussian-process regression and deep neural networks trained on phase-field or finite-element simulations as surrogates or direct predictors of corrosion curves and key electrochemical metrics [۱۳-۱۵]. The third domain covers composite systems such as PEO/layered double hydroxide (LDH) and nanoparticle-reinforced coatings on AZ-series Mg alloys, where ML links divalent cation chemistry, nanoparticle type and loading, and PEO conditions to combined corrosion and tribological performance [۶,۱۶]. Complementary work uses AI-based bibliometrics and deep-learning image segmentation to map research trends and quantify porosity and discharge traces in PEO layers [۷,۱۷]. Most available data sets are small and domain-specific, with limited sharing of raw data and feature definitions, which restricts model transferability between systems. Tree-based ensembles and shallow neural networks are the most widely adopted algorithms, while deep learning is mainly reserved for image-based or simulation-derived inputs [۳,۱۳-۱۵,۱۷]. Only a minority of studies employ explainable-AI techniques or multi-objective optimisation of coatings. We therefore call for better data curation and sharing, standardised physically meaningful descriptors, transparent modelling workflows, and further development of physics-informed and hybrid ML models combined with real-time monitoring and closed-loop control. Overall, this review maps current ML work in PEO and outlines steps towards more robust and transferable data-driven models for plasma-electrolytic surface engineering.

کلیدواژه ها:

artificial intelligence ، machine learning ، plasma electrolytic oxidation ، surface engineering