Training history and performance of the optimal neural networks composing the giant-impact model. Panels a–g: Mean squared error (MSE) between y(x) predicted by the neural networks g_i(i = 1,...,7) and the corresponding SPH outcome, t(x), as a function of the training epochs. The inset plots show the correlation with index R between y(x) and t(x) for the testing set. The complete training report (including the suboptimal neural networks) is in the Dataset S2.