In order to deal with the bending deformation of long span and heavy load crossbeams, a method is proposed to improve the geometric precision of the crossbeam, which combines FEA numerical computation and BP neural networks to predict the bending deformation curve of the crossbeam and prefabricate the compensation curve. The slide carriage is located at a series of working states, and the deformation is calculated using ANSYS software in order to obtain the training samples. By adjusting the appropriate parameters in Matlab software, the BP neural network is established to satisfy the error requirement. The deformation and compensation curve of the crossbeam are predicted through the trained neural network and the crossbeam is thus manufactured according to the compensation curve. The results of the bending deformation measurement demonstrate that the predicted values of the neural network match well with the experimental results, in which the relative error is less than 15% and the computing time is just 0.27 seconds. This method can accurately predict the bending deformation of the crossbeam and carry on compensation. Moreover, it provides new ideas for research as well as technical guidance for improving the crossbeam structure design and obtaining a perfect arch in advance.