Soil aggregates are the main soil components, in which carbon (C) and nitrogen (N) content and dynamics significantly influence the soil Cand Ncycle process, stability and soil fertility. Due to the difference of aggregates fractionation methods, the size of aggregate fractions obtained from different studies was not the same. Therefore, a large quantity of aggregates samples was required when using infrared spectroscopy to predict the properties of soil aggregates, while it is difficult to reasonably predict each fraction. Comprehensive modeling and prediction of samples from different aggregate fractions were conducted. Fourier-transform infrared spectroscopy analysis was carried out on the soil samples of the light chestnut soil in Inner Mongolia, using a genetic algorithm to select the characteristic wavelength. Prediction models of soil organic carbon (SOC) and total nitrogen (TN) in aggregate fractions were established based on partial least squares (PLSR), support vector machine (SVM), artificial neural network (ANN) and random forest (RF) methods. Based on the characteristic spectral interval screened by genetic algorithm, the ANN model showed the best modeling and prediction abilities of SOC and TN content in soil aggregates(RPD>2), which is significantly better than PLSR, SVM and RF models. The prediction ability of the ANN model based on full-spectrum data is lower than that of the ANN model based on GA-selected characteristic spectral intervals. The results indicated that the selection of GA-based characteristic spectral intervals could not only simplify the model structure and eliminate irrelevant information but also improve the accuracy and prediction ability of the model. In the present study, FTIR data from different aggregate fractions were mixed for modeling. By using a genetic algorithm to filter the characteristic spectrum, we found that the artificial neural network model can reliably predict the SOC and TN contents in soil aggregates, which was not affected by aggregate size.This might be mainly due to the fact that some wavelength ranges reflecting soil minerals, clay particles, etc. have been included in the selection of characteristic spectra by genetic algorithms, and that the effect of particle size on the SOC and TN might have already been included in the ANN model. The result highlights that the screening of characteristic wavelength intervals based on genetic algorithms and the use of artificial neural networks can model soil aggregates of different particle sizes in a unified manner, and can be used to estimate SOC and TN contents of aggregates.