Utilising a deep neural network as a surrogate model to approximate phenomenological models of a comminution circuit for faster simulations

•A neural network surrogate model is used to approximate phenomenological models.•The surrogate model only required 1 in 1000 data stratified sampling for training.•Evolutionary algorithm is used to optimise surrogate model hyperparameters.•The surrogate model is 3363 times faster compared to phenomenological models.•Low approximation error (0.37%, 0.55%, 0.45%) compared to phenomenological models.