To achieve rapid and accurate prediction of the natural carbonation depth of in-service bridge concrete under ordinary atmospheric environments, a dataset consisting of 496 groups of field-measured concrete carbonation samples from bridges is first established. Artificial Neural Network (ANN), Support Vector Machine (SVM) and Random Forest (RF) algorithms are adopted to screen and optimize feature combinations. Regression trees and four tree-based ensemble learning (EL) algorithms are separately applied to develop prediction models for the natural carbonation depth of bridge concrete, followed by interpretability-based machine learning analysis. The results show that: the semi-empirical semi-theoretical model specified in relevant current Chinese standard produces highly scattered predictions for natural carbonation depth of bridge concrete, while the EL model built with the Gradient Boosting Regression Tree (GBRT) algorithm delivers outstanding performance and enables fast and precise prediction results. Among the factors affecting natural carbonation of bridge concrete, concrete compressive strength and curing age stand out as the two most dominant features, negatively and positively correlated with carbonation depth respectively. The carbon dioxide concentration and annual average ambient temperature at the bridge site also exert considerable positive impacts on carbonation depth, while the remaining factors impose marginal effects.