Evolution of the mLAE on the test set over time across surrogates and datasets. To investigate error accumulation over time, we compared the one-shot mean, where predictions are obtained by providing the initial state and the desired time step, with the iterative mean. For the latter, we obtained predictions by dividing the domain into intervals of i = 10 time steps (indicated by the vertical dashed lines), and using the predicted abundances at the end of one interval as initial state for the next interval. In this iterative setting, fully connected surrogates generally exhibit a more pronounced accrual of error over time, while latent-evolution models prove more robust.