In order to overcome the dilemma of insufficient effective sample data for subsurface microcrack of brittle materials with fixed abrasives and achieve effective prediction of subsurface microcrack depth, a co-training support vector regression algorithm (Co-training SVR) was used to construct the prediction model. The effects of different labeled training set partitioning methods on the mean square error of test set were compared. Then the predictive performance of supervised learning PSO-SVR was compared with that of the model. Finally, brittle materials such as glass-ceramics and calcium fluoride, which were not included in the labeled training set, were taken as processing objects for lapping and angular polishing experiments, and four groups of subsurface microcrack depth values were compared with the predicted values of the model. The results show that co-training SVR model with separate partitioning method have smaller convergence value of mean square error. Compared with the PSO-SVR model, the mean square error and mean absolute percentage error of the proposed model are reduced by about 9% and 17%, respectively. The prediction error of the model for the four groups of verification experiments is between 1.2% and 13.8%. The above results show that the model can predict the subsurface microcrack depth accurately and reliably when lapping brittle materials with fixed abrasives.