This paper reports on a general approach to build a large-signal, neural network HEMT model using a genetic algorithm. By representing the configuration of a neural network model as the chromosome of a virtual creature, we looked for an optimum network configuration by simulating the evolution of a group of these virtual creatures (a population). We successfully designed neural networks representing bias-dependent intrinsic elements of a HEMT's equivalent circuit. We also verified the reliability of this technique by searching for the optimum model from different initial conditions.