Note that even though currently multi-objective tuning is not supported in mlr3, there are similar situations where these multi-objective problems appear and pareto optimal solutions that represent the best trade-off between (2 or more) objectives have been proposed (as there are many algorithms on findings "knees" of the (multi-dimensional) pareto front - see a nice review in this article).

In a recent feature selection example, I implemented a very simple 2d knee point identification method, to find the pareto point point with as minimum selected features as possible, while retaining as high performance as possible, see this mlr3 gallery post.