MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields

We propose that these limitations arise because MPNNs only pass two-body messages leading to a direct relationship between the number of layers and the expressivity of the network.