You need a network with at least two hidden layers to model 1/a, since the function is nonlinear. If a is in [-1, 1] interval, the function is a bit hard to model since it jumps from -infinity (for negative values) to +infinity (for positive values), i.e. it is not continuous. In the first hidden layer, I would put 2 neurons (with ReLU activations) to model the cases a < 0 and a > 0 independently. In the second hidden layer you need more neurons (try 10 neurons with logistic or tanh activations) to model the nonlinear 1/a function. Maybe this blog post can provide some additional insight: https://towardsdatascience.com/can-neural-networks-really-learn-any-function-65e106617fc6
If a is in (0, +infinity), the function is indeed continuous. In this case, you just need to use enough neurons to model the function with the desired accuracy (below a certain epsilon).
When switching to machine learning powered by Clifford/geometric algebra, simple operations like summation, division, multiplication and square root are well-defined geometric operations.