On the impact of larger batch size in the training of Physics Informed Neural Networks