This is the sort of problem that some astute design calculations should be able to disentangle. ;)

Maybe they were using a topology optimizer driven by reinforcement learning to come up with the initial Pareto front.