Neural networks have provided a promising avenue of research for predicting the states of many-body wave functions. However, it is computationally expensive to confirm the accuracy of the model for the time evolution, resulting in slow or intractable accuracy checks during the training process for larger systems. Furthermore, the most commonly used methods rely on inputs of fixed sized, resulting in models that cannot be applied to systems of differing dimensions. In this thesis, we test three different methods of transfer learning in an attempt to resolve these issues. The first, a tiling method for a restricted Boltzmann machine, resulted in convergence in fewer iterations, though did not significantly improve accuracy. For convolutional neural networks, the traditional approach of freezing layers did not result in significantly improved performance, although use of the reptile domain generalization algorithm gave similar results to the base models, while allowing for varying input dimensions. No method seemed to significantly increase the accuracy of time evolution. Domain generalization may provide a way to create models that are not as rigid with regard to input, while other architectures may provide better performance for other tasks.