Efficient training of artificial neural network surrogates for a collisional-radiative model through adaptive parameter space sampling