Schematic representation of the domain-adaptation problem tackled in this work. The task consists in inferring the classification of PHANGS nebulae, given a set of photoionisation simulations from nebulae of different classes. The source domain (theory) contains the labels, while the target domain (data) is unlabelled. A classical, supervised machine-learning algorithm (e.g. a classifier neural network, in blue) is trained on the instances and labels of the source domain, but may perform poorly on the target domain, because of the differences between simulated and real data. In this work we employed a DANN, a domain-adaptation algorithm that uses as input (in green) both the labelled source domain and the unlabelled target domain instances. This approach allows the model to learn domain-invariant features and perform better on the target domain.