Carefully constructed models on what should exist, and how those things should behave if they exist, then experiments that show these things behaving as predicted, therefore validating the model. It's an inductive process like dialectics. You move between theory and praxis, updating theory and adjusting action. Though this kind of paints a picture that all science is experimental when it's not and yet is equally valid. Like lots of stuff in geology can't be observed directly due to time scales or ran as a experiment. In a way there's more experimental data supporting the wave-particle duality of light than experiments supporting plate tectonics. It becomes just thinking about stuff really hard, and looking at stuff very closely to make sure you're observing it the right way. Make predictions about what you should find next if you're correct.