Modelling of nuclide densities as a function of time within magnetic confinement fusion devices such as the JET, ITER and proposed DEMO tokamaks is performed using Monte-Carlo transport codes coupled with a Bateman equation solver. The generation of reaction rates occurs either through point-wise interpolation of energy dependent tracked particle data with nuclear data or multi-group convolution of `binned' fluxes with binned cross-sections. The multi-group approach benefits from decreased computational expense, but introduces errors through effects such as self-shielding. Depending on the multi-group structure and nuclear data used, this method can introduce unacceptable errors without warning. We present a multi-group optimisation method which utilises a modified particle swarm algorithm followed by a non-stochastic `string-tightening' algorithm. This procedure has been used with a semi-homogenised 1D DEMO-like reactor design in order to produce an optimised energy group structure for tritium breeding. In this example, the errors introduced by the Vitamin-J 175 multi-group are reduced by two orders-of-magnitude in the optimised group structure.