Early-stage forecasting of bronze disease development with chlorine mapping: Integrating computer vision and multimodal characterization methodologies

To advance non-destructive evaluation of bronze artifacts, this work presents an approach integrating multimodal data from corrosion images and MA-XRF chlorine distribution maps for characterizing corrosion products induced by bronze disease. Dry-wet cyclic experiments were conducted to simulate corrosion processes, from which corrosion images at different time points were systematically acquired. Chlorine elemental distribution maps were obtained via macro X-ray fluorescence (MA-XRF) imaging, from which the area proportion of high-chlorine regions was extracted for data mining, modeling, and experimental validation. Results show that 20 physical features extracted from corrosion images exhibit significant Spearman correlations (ρ > 0.4, up to 0.6) with high-chlorine area fractions, validating the feasibility of inferring bronze disease progression from visual characteristics. Machine learning models, trained on these visual features to predict chloride-rich area fractions, achieved an R² of 0.83, demonstrating robust capability for forecasting bronze disease evolution directly from images. A clustering-based classification model, integrating multi-modal physical features, categorizes corrosion products into four distinct classes, elucidating the spatiotemporal dynamics of rust layer transitions in the early stages of bronze disease development. This approach enables a preliminary assessment of the progression of bronze disease.