Fat plays an essential role in the composition of infant formula. Not only is fat a vital component of a babys growth and development, but it also provides essential energy for growth. It is crucial for the development of the infant brain and the formation of nerve myelin. Chemical methods for determining the fat content of infant milk powder, such as ether extraction, are sensitive but have the disadvantage of destroying samples and having a long detection period. In this paper, the hyperspectral data undergoes preprocessing processes with standard normal transform (SNV), multiple scattering corrections (MSC), Savitzky-Golay smoothing, and Roust method using different stages of infant milk powder in Inner Mongolia, China. A competitive adaptive re-weighting algorithm, CARS, was used to sift out redundant wavelengths from the spectroscopic data at 125 feature wavelengths, leaving 66 valid wavelengths. The Bayesian optimization algorithm optimizes the XGBoosting prediction model, leading to a BO-XGBoosting model that predicts the fat content of infant formula better than the original model. The experimental results show that the model predicts better than the traditional partial least squares regression (PLSR) and support vector machine (SVR) regression model, outperforming the Bagging and GrdientBoosting algorithms in the integrated algorithm. In the BO-XGBoosting model in the test set experiments, the decision coefficient R2 and root mean square error of prediction (RMSEP) obtained are 0. 953 7 and 0. 577 3, which are 2. 91% higher and 19. 2% lower than the determination coefficient R2 and root mean squared error of prediction (RMSEP) of the XGBoosting models R2 and RMSEP, respectively. This study provides algorithmic support and a theoretical foundation for BO-XGBooting based rapid, non-destructive detection of infant formula fat content.