I mean, use the ML model to predict where they will be and then do an experiment to confirm it. My understanding from what I've read is that it's easier to make sense of experimental data (and likely cheaper) when you have a good idea what you're looking for.