Microstructure-agnostic deep learning for mechanistic discovery of corrosion-resistant Co-Cr-Fe-Ni MPEAs

In multi-principal element alloys (MPEAs), the vast compositional design space and limited availability of microstructure-resolved corrosion data pose significant challenges to the predictive design of corrosion-resistant compositions. Traditional approaches rely heavily on trial-and-error experimentation and detailed microstructural characterization, which are time-consuming and resource-intensive. In this work, we develop a microstructure-agnostic deep learning framework that predicts the corrosion resistance of Co-Cr-Fe-Ni MPEAs directly from composition-based descriptors. By integrating physics-informed features, active learning, and uncertainty quantification, the model captures key physicochemical trends without requiring explicit structural input. The framework rapidly identifies non-equiatomic compositions with superior corrosion resistance, validated experimentally to outperform 304 L stainless steel under both acidic and chloride-rich conditions. Fundamental insights into corrosion mechanisms were obtained by linking model-predicted corrosion trends with electronic structure properties from density functional theory (DFT). The most corrosion-resistant compositions—e.g., Fe34Cr34 and Co34Cr34—exhibit high electron work functions and oxygen binding energies, indicative of strong passivation tendencies. Atom probe tomography (APT) of corroded surfaces confirms that these alloys form thinner (∼9 nm) and more protective passive films compared to conventional Cr-rich MPEAs (∼14 nm). Feature importance analysis within the machine learning model aligns with DFT-calculated elemental contributions to surface energetics, revealing a consistent ranking of Cr > Fe > Co > Ni in promoting passivation.