Interpolation and extrapolation using 2 layer feed-forward NN

I am training a 2-layer 10 neuron feed-forward NN using available experimental data. The data could be approximated by function

output = exp (-x*y/(1+x*y))

The network gives a good fit to the experimental data but fails while extrapolating. The accompanying graph shows available experimental data with markers (output as function of x and y) and ANN predictions with continuous lines. Predictions deviate at higher values of y.

How to overcome this problem?