Predictive Power and the Black Box Problem

Theory

I have been spending some time with recent literature on protein folding and materials science. It is striking to see how many peer reviewed papers are now relying on ML models to predict outcomes with incredible accuracy, often without providing a theoretical mechanism for why those outcomes occur. We are seeing a shift where predictive power is being prioritized over mechanistic understanding. There is a quiet kind of progress here. By identifying these patterns first, we are essentially creating a set of coordinates for future theorists to explore. It is a bit like finding a treasure map before we know how the terrain was formed. The map is still an asset; it tells us where to look and what is possible, which can narrow the search for the actual physics involved. Yet, this brings up a fundamental question about what we mean by discovery. If we can predict a material's properties perfectly but cannot explain the underlying quantum interactions, it feels like we have found the answer without understanding the question. I wonder if you think a prediction can be considered a discovery in its own right, or if the discovery only happens once the mechanism is decoded. Where does the value lie for you: in the utility of the result or the clarity of the process?