To add to the above, it's the difference between telling a model:
"Model, show me spectral signatures which are similar."
And "Model, is there anything interesting about the spectral signatures collected?"
In the former, you are potentially able to (or your NN layers can) identify representative features for your intended result.
In the latter, you're asking machine learning to synthesize the sum of human scientific knowledge, then extract interesting facts from the data using it. E.g. "Hmm, the vortex patterns on this Jupiter storm are incongruous with our fluid dynamics models" or "The reflectivity of the surface seems to indicate a different composition than we expected."
AI isn't currently capable of formulating questions, asking them, answering them, and ranking the results on significance.
And in lieu of that, the only use is "Help a scientist answer a question they posed." Which seems to be the initial problem that started this thread! Not enough scientists can get grants to pose all the questions that should be asked of old data.