One of the key performance metrics for optical networks is the maximum achievable throughput for a given network. Determining it however, is an NP-hard optimisation problem, often solved via computationally expensive integer linear programming (ILP) formulations, infeasible to implement as objectives, even on very small node scales of a few tens of nodes. Alternatively heuristics are used, although these too require considerable computation time for large numbers of networks. There is, thus, a need for ultra-fast and accurate performance evaluation of optical networks. For the first time, we propose the use of a geometric deep learning model, message passing neural networks (MPNN), to learn the relationship between, node and edge features, the structure and the maximum achievable throughput of networks. We demonstrate that MPNNs can accurately predict the maximum achievable throughput while reducing the computational time by up to 5-orders of magnitude compared to the ILP for small networks (10-15 nodes) and compared to the heuristic for large networks (25-100 nodes) -proving their suitability for the design and optimisation of optical networks on different time-and distance-scales.