Performance evaluation of machine learning techniques in surface morphology and corrosion prediction for A286 3D printed micro-lattice structures.
Journal:
PloS one
PMID:
40334230
Abstract
The development of lightweight, corrosion-resistant metallic lattice structures has gained significant attention in aerospace, defense, and structural applications, where material durability and weight optimization are critical. This study investigates the corrosion behavior of Laser Powder Bed Fusion (LPBF)-fabricated A286 steel honeycomb, Body-Centered Cubic (BCC), and gyroid lattices, comparing their performance against conventional materials such as Rolled Homogeneous Armor (RHA), Maraging High Strength Steel (MHA), and High-Nitrogen Steel (HNS). Corrosion testing was conducted using accelerated salt spray exposure, and the results were validated through computed tomography (CT)-based structural integrity analysis and machine learning-based predictive modeling. The experimental findings revealed that lattice structures exhibited significantly lower corrosion rates than conventional bulk materials, with the honeycomb lattice demonstrating the highest corrosion resistance (1.218 mm/year), followed by BCC (1.311 mm/year) and gyroid (1.671 mm/year). Compared to RHA, the honeycomb lattice exhibited a 57.23% reduction in corrosion rate, confirming its superior electrochemical stability. CT scan evaluations further highlighted differences in density distribution and geometric fidelity, with honeycomb lattices showing the most uniform porosity, while BCC structures displayed localized density variations at nodal intersections. To enhance predictive capabilities, various machine learning (ML) algorithms were employed to model corrosion behavior based on weight-loss measurements and lattice topology. Bayesian Ridge regression outperformed other models, achieving an R² of 0.99849 and RMSE of 0.00049, confirming its robustness in capturing corrosion trends. Linear Regression also performed well, while ensemble models such as Random Forest and XGBoost exhibited higher error margins due to dataset linearity constraints. Residual analysis and graphical interpretations further validated the stability and predictive reliability of ML-based corrosion assessments, demonstrating their feasibility as an alternative to traditional experimental methods. This study presents a comprehensive framework for integrating experimental corrosion testing, computational modeling, and CT-based defect analysis, offering a scalable approach to optimizing micro-lattice designs for corrosion-sensitive applications. The findings highlight the potential of LPBF-fabricated metallic lattices for aerospace, defense, and marine structures, where enhanced corrosion resistance, reduced material degradation, and predictive maintenance strategies are essential for long-term operational performance.