Improving the performance of Stein variational inference through extreme sparsification of physically-constrained neural network models