Design of rust layer stabilizers for weathering steel guided by interpretable machine learning and Bayesian optimization

Abstract

Surface stabilization treatment serves as a primary method to promote stable rust layer formation on weathering steel (WS). However, due to the complex and multicomponent chemical formulations of stabilization treatment agents (STA), the precise control over STA component ratios to achieve the best stabilization treatment effect remains highly challenging. This study combines high-throughput experiment and machine learning method to establish an optimization framework for designing rust layer STA formulation. By employing high-throughput droplet dispensing experiments and wire beam electrode electrochemical testing, a predictive model is constructed using the AdaBoost algorithm. Interpretability analysis is further integrated to guide Bayesian optimization for iterative formulation refinement. After two optimization cycles, the optimal STA formulation (0.70 g/L CuSO₄, 0.20 g/L MgSO₄, 0.60 g/L Na₂HPO₄, and 0.20 g/L tannic acid) is identified from over 2.8 million candidate formulations. The optimized STA promotes the generation of stable rust layer on Q420 WS, which effectively reduces rust layer defects, inhibits corrosive medium penetration, and significantly enhances the corrosion resistance of WS.