We used artificial neural networks to replace the complete spin-up procedure that computes a steady annual cycle of a marine ecosystem driven by ocean transport. The networks took only the few biogeochemical model parameters and attempted to predict the spatially distributed concentrations of the ecosystem, in this case only nutrients, for one time point of the annual cycle. The ocean circulation was fixed for all parameters. Different network topologies, sparse networks, and hyperparameter optimization using a genetic algorithm were used. This showed that all studied networks can produce a distribution that is point-wise close to the original spin-up result. However, these predictions were far from being annually periodic, such that a subsequent spin-up was necessary. In this way, the overall runtime of the spin-up could be reduced by 13% on average. It is debatable whether this procedure is useful for the generation of initial values, or whether simpler methods can achieve faster convergence.