A Helmholtz equation solver using unsupervised learning: Application to transcranial ultrasound

•We train a neural network to solve the heterogeneous Helmholtz equation.•The algorithm requires considerably less iterations than GMRES.•Using a physics loss, no ground-truth data are needed.•The average maximum error to a reference solution for the test set is 0.36%.•The network is capable of generalizing on problems outside the training set.